AN ABSTRACT OF THE DISSERTATION OF

Andrew M. Abate for the degree of Doctor of Philosophy in Robotics presented on

April 11, 2018.

Title: Mechanical Design for Robot Locomotion.

Abstract approved:

Jonathan W. Hurst Ross L. Hatton

Robotic limbs have been shown to enable mobility in unstructured, real-world terrain;
they allow robots to step around cluttered environments, scramble up hills, carry heavy
loads, and even perform acrobatics. However, mechanical limbs cannot operate as a means
for such dynamic locomotion if they are simply treated as general articulated mechanisms.
Mechanical design in the context of locomotion requires us to be aware of the control
paradigm (sometimes even serving as a part of the controller) and how the robot relies on
contact to generate motion. These two interfaces (software control, physical contact) direct
the design of inherent hardware dynamics (passive dynamics), where mechanical design
decisions are made specifically to serve the context. In this work, we create a practical
definition of locomotion to lead us towards metrics for grading a design’s aptitude for
real-world mobility and tracking progress toward better designs. Metrics such as weight,
cost, and other engineering goals are commonly understood and not considered here.
Instead, we are interested in the locomotion-specific metrics of floating-base
manipulability, impact reduction, power quality, and compliance tuning. These metrics
have intuitive, geometric representations for use in human-driven design processes, while
their numerical representations can be evaluated in simulation. Multiple competing
metrics require a multi-objective design process, which here is discussed as stochastic
search. This structure is analogous to a human-driven design process, but is also amenable
to future computer automation.
© Copyright by Andrew M. Abate
April 11, 2018
All Rights Reserved
Mechanical Design for Robot Locomotion

by
Andrew M. Abate

A DISSERTATION

submitted to

Oregon State University

in partial fulfillment of
the requirements for the
degree of

Doctor of Philosophy

Presented April 11, 2018

Commencement June 2018
Doctor of Philosophy dissertation of Andrew M. Abate presented on April 11, 2018.

APPROVED:

Co-Major Professor, representing Robotics

Co-Major Professor, representing Robotics

Head of the School of Mechanical, Industrial, and Manufacturing Engineering

Dean of the Graduate School

I understand that my dissertation will become part of the permanent collection of Oregon
State University libraries. My signature below authorizes release of my dissertation to any
reader upon request.

Andrew M. Abate, Author

 TABLE OF CONTENTS

Page
1 introduction 1
1.1 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 locomotion systems 7
2.1 The first walking robots, static stability . . . . . . . . . . . . . . . . . . . . . 8
2.2 Dynamic stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Passive dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Active dynamics through high-transparency actuators . . . . . . . . . . . . 12
2.5 ATRIAS and Cassie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 modeling virtual prototypes 20
3.1 Software, hardware, and the world . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Core mechanical system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Control interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4 World interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4 defining the locomotion task 34
4.1 Abstract locomotion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Spring-mass locomotion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Through-contact trajectory optimization . . . . . . . . . . . . . . . . . . . . . 40
5 performance metrics 41
5.1 Efficiency metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2 Manipulability metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.3 Robustness metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.4 Feasibility metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6 metric-driven design process 62
6.1 Automated design generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7 example designs 66
7.1 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
8 conclusion 76
8.1 Disclosure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
8.2 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

bibliography 78
 LIST OF FIGURES

Figure Page
1 Limbed locomotion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Candidate locomotor design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3 PV-II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4 Humanoids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
5 Leg design silhouettes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
6 Passive-dynamic walking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
7 IHMC Elliptical Runner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
8 ATRIAS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
9 Cassie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
10 Cassie leg design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
11 ATRIAS motor power requirements . . . . . . . . . . . . . . . . . . . . . . . 19
12 Three dynamic subsystems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
13 Multibody system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
14 Plate-spring models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
15 Spring-loaded inverted pendulum model . . . . . . . . . . . . . . . . . . . . 37
16 Antagonistic power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
17 Power quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
18 Power quality in three leg designs . . . . . . . . . . . . . . . . . . . . . . . . 48
19 Force, velocity decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
20 Compliance dyads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
21 Acceleration limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
22 Impact inertia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
23 Quadratic loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
24 Corner sort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
25 Design iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
26 Polar design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
27 Serial design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
28 Parallel design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
“In design, one of the most difficult activities
is to get the specifications right.”
— Donald A. Norman, The Design of Everyday Things
1 INTRODUCTION

We want to improve robot mobility. Wheels and tracks are good for traversing simple,
smooth, and open environments, but limbs are the key to robust terrestrial locomotion
(Figure 1). The seemingly simple use of limbs for self-transport is a gating issue for a
stampede of applications: helper robots in homes, carrying bags, and moving furniture;
courier robots on the streets delivering packages; search and rescue robots in forests and
caves assisting disaster response; geological robots climbing the mountains of Earth or
Mars to survey the landscape; and so on as far as our imagination takes us. Such activities
seem trivial to us, but locomotion is a deceptively complex problem.
Dynamic locomotion is as much a hardware design problem as it is one of control.
Studies in control theory abound, but there is a relative fuzziness in the controller’s
companion: a purposeful mechanical design. The algorithms and processes which make
up a controller cannot directly act on the environment, so they need a vehicle or transducer
to act through. Hardware must always serve that function by providing a controllable
and efficient platform for locomotion. Clearly, a locomotor will be underwhelming if its
hardware dynamics are too slow or heavy for walking and running. Conversely, a nimbler

Figure 1: Limbed robots have the potential for robust terrestrial locomotion in diverse settings;
they can fold to squeeze through tight spaces, extend to reach ledges, hurdle over or
bridge gaps in the environment, and use arms for both locomotion and manipulation. A
wheeled or tracked robot is limited by its fixed size, either being too large to fit through a
space or too small to get over an obstacle [82].

1
 introduction 2

hardware design could increase controllability and even manage real-world events (e.g.
impacts) that are above the controller’s bandwidth. It is even possible for mechanical
design to affect how complex the control system must be (e.g. Cartesian kinematics make
IK/FK trivial [58], gibbons effortlessly swing through natural pendular trajectories [17,
147], dead fish can "swim" upstream because of their bodies’ passive dynamics [16], etc).
These facts combined with the fixed nature of a hardware design makes mechanism
dynamics critical to "get right"; software can be patched, but it is certainly challenging to
hotfix deployed machines.
So, we have a big question to tease apart: "how can we make better locomotion hard-
ware?". As humans, we have an innate sense of animals walking and running, but what
does "locomotion" mean for a robot? Does better performance depend more on control,
hardware dynamics, or some balance of the two? How do we even define better? We
sometimes have clear metrics for better when it comes to making the robot lighter, or
selecting actuators with higher torque density, but we are quite often forced to make
judgment calls† when confronted with the complexity of a locomotion system. Does the
robot work better with forward-facing knees or backward, and why? Is it important where
actuators are located in the robot, or do the limbs simply need to span the workspace?
How much do we need to know about the control system before designing the hardware
(e.g. [92])? Is physical compliance good or bad, or both? With enough effort, we can answer
these questions and find better designs by using a formal design process.
The goal of this work is to better understand how the demands of dynamic loco-
motion influence a robot’s mechanical design, and how to manage design goals and
constraints in a methodical design process. In order to improve designs, we must have
a rigorous structure for ideation, evaluation, and iteration. This repeated generation and
ranking of designs can be formalized as an optimization problem. Optimization is perhaps
the wrong word due to the levity of human thought and preference as compared to a
computer, but the format of an optimization problem gives us the perfect structure for our

† Judgement calls in design can be troublesome; people have different "gut instincts" about design, sometimes
leading a project around in circles. A project needs to decide on the appropriate metrics for success in order to
make structured progress.
 introduction 3

design problem and forces us to be specific about our design goals. It also tempts us with
the possibility of automating some or all of the process of finding better designs.
If the chosen metrics are insufficient or direct the search in an undesired direction, then
the metrics need to be re-evaluated. That being the case, choosing metrics is just as much a
problem as choosing a design. We do not have a list of design objectives which definitively
produce optimal designs, but getting closer to that ideal is a main driving force behind
this work.

1.0.1 Formulating the design problem

Let us cast the locomotor design problem into the same three parts as an optimization
problem, as these will guide us in teasing apart the question of designing better machines
for locomotion:

• A design space of all the possible candidate robots. This is all the different configu-
rations of links, masses, actuators, and springs. However nonsensical some designs
may appear to us, all possible designs are necessarily in the search space.

• A set of design objectives allow candidate designs to be ranked against each other.
Some objectives may apply strictly to the structure and composition of the robot, and
others may also require a task definition such as a walking trajectory. All objectives
discussed in this work relate specifically to locomotion. Engineering objectives such
as component mass and cost must obviously be considered, but they are intentionally
avoided in favor of locomotion metrics.

• Constraints ensure the feasibility of designs. These may apply at the design level
where links cannot have negative mass or be shorter than a given length, or they can
enforce the peak torques and speeds of actuators, or otherwise make sure that the
robot is physically capable of following a task’s trajectory. Engineering constraints
such as component failure and fatigue are not discussed, but obviously must be
included.
 introduction 4

Design Objective Score Order

Mass 30 kg <
CoM acceleration limit 2.1 g >
Toe acceleration limit 25 g >
Nominal compliance 11 kN/m =
Axial compliance + (-,~,+)
Impact inertia 2.5 kg <
Power quality ~ (-,~,+)
Physics violation _ (-,~,+)
... ... ...
... ... ...

Figure 2: Candidate designs are members of the search space and are evaluated according to the
project’s objectives. Objectives may or may not carry with them extra information about
the use case or the task, such as trajectories, controllers, or environments.

Each design candidate is a member of the design space that has been evaluated according
to the design objectives (Figure 2). Depending on the objectives, each candidate may need
to include extra data to describe the task. For example, candidates may be evaluated
relative to some common reference trajectory, reference controller, or an optimization
which determines optimal trajectories for each new candidate.

1.0.2 Solving the design problem

Given the combinatorial structure of the search space, the locomotor design process is
naturally iterative [26, 72]. We cannot deterministically arrive at the ideal design, so
successive candidates must build upon earlier successes and failures. We also cannot know
when a solution is truly optimal, but we can easily check that some identified solutions
are better than others. So, the only way to solve the design problem is to continually find
better designs. We can never find the ultimate design.
In fact, there is no ultimate design, because having multiple competing objectives means
that designers must make trade-offs depending on the nature and scale of their project. The
existence of trade-offs means that we are inherently searching for an optimal set of designs.
It is up to the designer to decide what trade-offs to make and which design to pursue.
Designs can either be hand-picked from the set of current best designs, or designers can
 1.1 contribution 5

weight different metrics to give each design a single combined score with a single winner
[6].
We cannot evaluate designs without first generating them, so we need a way to instantiate
particular designs from the search space. The design space we consider in this work
is all the different assemblages of rigid bodies, joints, actuators, and passive springs.
Algorithmically searching over such a space requires a formal grammar (production rules)
for strings describing valid members. Fortunately, humans are extremely good at idea
generation, so we leave it up to them for the time being. The only downside of ideation is
ensuring good coverage of the search space, but generating all strings from a grammar
can be equally difficult (depending on the type of grammar). Regardless, this is precisely
why we have industrial designers and engineers.

1.1 contribution

This work is the result of trying to create better hardware for locomotion systems. It
investigates how a robot’s mechanical design is influenced by the demands of locomotion,
and determines logical design metrics and goals.
I have led the mechanical design of two bipeds at Oregon State University: ATRIAS 2.0/2.1
followed by three revisions of the Cassie biped† . It took a substantial amount of time to
understand the specialized mechanical requirements of locomotors and how connected
the hardware is to the control system, so I aim to render that knowledge here.
I have developed the use of task-space compliance and inertia in design [2], and I have
derived the Power Quality metric to maximize mechanical efficiency under uncertainty
[3]. These metrics were used in the Cassie design to minimize the amount of power lost to
motor braking while maintaining good passive dynamics.
Wherever possible, I translate numerical metrics into tangible, geometric forms which
provide intuition about candidate designs. Visual forms are often self-explanatory, easily
transferring design intuition while also being backed by concrete mathematics.

† The Cassie biped design is licensed to Agility Robotics for production and sale. Cassie robots are being sold
to universities working to advance legged locomotion capability.
 1.2 outline 6

1.2 outline

locomotion systems Previous examples of locomotion systems show the breadth

of thought and approaches to locomotion.

modeling virtual prototypes It is not feasible to build and test all candidate
designs in the real world, so we must sometimes simulate virtual prototypes. There is
substantial modeling effort required to simulate a mechanical system interacting with the
world through contact.

defining the locomotion task We need a clear definition of what the robot
is trying to accomplish before we can create a list of metrics to rank candidate designs.
Locomotion has both abstract and concrete task definitions in that both "system self-
transport" and "left-right-left-right" walking are locomotion. Both of these definitions are
good for different purposes.

performance metrics With the task definitions, we can arrive at logical design
goals which specifically target locomotion performance. We consider performance in terms
of energy efficiency, manipulability, robustness, and feasibility.

metric-driven design process Finally, the virtual modeling, task definition, and
objectives are collected into an iterative design process which is most closely described as
stochastic search. Such a formalism is analogous to the engineering design process, and it
is also amenable to computer automation.

example designs We show how limb-design metrics can be represented visually to

judge candidate designs. We evaluate the acceleration limit, impact inertia, and compliant
response of a polar design, a serial design, and a parallel design.
 2 LOCOMOTION SYSTEMS

Researchers, engineers, and artists have been trying to make walking machines for at
least a century‡ [137], and we are only beginning to see glimpses of natural agility and
efficiency. The robots with the most natural movement tend to be those which have been
focused on hardware designed specifically for locomotion [63, 93, 115]. Designs which do
not consider their passive response to contact tend to be the most "robotic", attempting to
maintain static equilibrium and move slowly to avoid exciting unwanted dynamics (e.g.
humanoids like ASIMO and Valkyrie).
Since the early 1980s, marking the beginning of using computers to drive legged
machines, the robotics community has created a substantial number of robots designed for
locomotion [77]. Anthropomorphic, "humanoid" robots compose the majority of the catalog,
while there are also a number of monopod hopping machines [65] designed principally
for research in dynamic locomotion and stability as well as bi-, quad- [66, 119], and
hexapods [129]. Most of the early robots, and especially humanoids, use rigid gearmotors
or hydraulics for actuation. Since the invention of the Series Elastic Actuator (SEA) [114,
126] and other compliant actuators [56], we have also seen a number of compliant leg
designs. More recently, we have seen a number of smaller robots eschewing gearmotors in
favor of high-transparency, direct-drive electric torquer motors. There are also a number
of machines designed to rely on hardware dynamics for stability, using only minimal (or
zero) actuation or control.

‡ See Reuben Hoggett’s wonderful collection of early walking machines at http://cyberneticzoo.com/

walking-machine-time-line/

7
 2.1 the first walking robots, static stability 8

Figure 3: The 1978 PV-II "daddy long-legs" robot and its Cartesian pantograph leg mechanism [58].

2.1 the first walking robots, static stability

Early walking machines in the 1980s were designed to accommodate simple theories of
balance based on static stability. Robots such as McGhee’s hexapod [95, 107] used a wide
base of support for better stability, and many humanoids to this day still have exaggerated
foot size for better area of support. This approach is found in Zero-Moment Point gaits
[150], where the robot moves slowly to reduce dynamic effects and ensures its mass-center
is always over its area of support.
Early robots had to contend with the lack of available computing power, so the PV-
II robot used a Cartesian leg design to mechanically reduce the cost associated with
calculating kinematics [58]. The PANTOMECH legs used on the PV-II move in 3 orthogonal
directions aligned with the robot’s Cartesian frame, making forward and inverse kinematics
trivial (Figure 3). This is a great example of how a robot’s hardware can be designed
specifically to aid a controller.

2.1.1 Humanoids, Static stability

Humanoid robots tend to be built as platforms for general control and manipulation. They
have predominantly rigid actuation and follow precise motion trajectories, much like their
fixed-base manipulator ancestors. Rigid humanoids have high acceleration potential when
 2.2 dynamic stability 9

they are in contact with rigid ground, and they can follow undisturbed trajectories, but
they tend to be delicate and have small basins of attraction (Figure 4).
Rigid humanoids are typically designed with enough arm, leg, and body actuators to
remain statically stable and interact with the environment. This allows them to be fully
actuated and fully controllable (within bounds), but the robots must have detailed maps of
the environment to avoid collisions which cause large force spikes and potentially damage
their rigid gearing. Large, uncontrollable impulses also make rigid humanoids sensitive to
disturbances. In order to get around these deficits, humanoids typically have large foot
pads to deaden impacts, and they are given a wide berth so as to not accidentally cause
unmodeled contact events.
One outlier is the ATLAS robot (Figure 4), which uses high-power hydraulics for more
robust locomotion. Driven by a number of competing teams including MIT and IHMC at
the DARPA Robotics Challenge in 2015, the ATLAS robot had some very real success (and
also a lot of failures, resulting in a gag reel watched more times than any other video of
the event). Using an ATLAS torso, Boston Dynamics’ Handle robot even has wheels for
feet (because, why not?).
SRI’s DURUS humanoid (also at the DRC as a dynamic demo) uses series-elastic ankles
for efficient locomotion [120], and IIT’s Coman [146] uses series-elastic actuators in most
of its joints.

2.2 dynamic stability

Static stability is limited to known, stable terrain, so robots needed to begin relying on
dynamic stability to make it into the real world. The idea of dynamic stability allowed
mono-, bi-, and quadruped robots to confidently move through an unknown and dynamic
world, reject heavy disturbances, and even do acrobatics while remaining stable [117, 118].
Marc Raibert’s machines as well as many other hoppers [5, 67, 161] and bipeds [63] are able
to move robustly with the combination of passive compliance and intuitive control laws
(such as placing the next step farther forward for higher speeds and injecting energy with
 2.3 passive dynamics 10

Figure 4: NASA’s Valkyrie, Honda’s ASIMO, and Boston Dynamics’ ATLAS.

each stride). Intuitive controllers inspired by Raibert’s work have enabled the compliant
MABEL [111], ATRIAS, and Cassie bipeds to walk and run at Oregon State University.
These and other machines try to embody reduced-order dynamics in order to simplify
control [47].
Other early hoppers include Brown’s Bow Leg Hopper [23], Andrews’ hopper [12], and
the Papantoniou straight-line linkage hopper [109] (Figure 5). Like Raibert’s hoppers, these
are designed with a low leg-to-torso inertia ratio and leg angle axes near the CoM for
simplified dynamics. As such, they are amenable to intuitive control schemes and are used
to test new theories in dynamic balance based on the spring-loaded inverted pendulum
(SLIP) and other reduced-order models.

2.3 passive dynamics

Taking the idea of optimized kinematics/dynamics to the extreme, passive-dynamic walk-

ers and runners can generate gaits without any control, electronics, or motors whatsoever
(Figure 6). Tad McGeer created a fully-passive mechanism — a pair of legs — which could
"walk" down a slight slope, powered only by gravity [94]. Since they lack any control or
actuation, these mechanisms are very sensitive to disturbances and cannot walk over level
 2.3 passive dynamics 11

Figure 5: Silhouettes of various compliant leg designs. From top left: bow-leg hopper, Andrew’s
hopper, ARL monopod, ATRIAS, MABEL, Raibert’s hopper.

ground [31, 32]. However, McGeer’s walkers evolved into minimally-actuated robots which
can walk on flat ground with astounding efficiency. For example, Cornell’s Ranger robot
was able to walk an Ultra-Marathon on a single battery charge [18]. So, even with access
to high-powered mobile computers, designing a robot’s hardware for a specific purpose
can still have huge advantages.
Other examples of designing for passive dynamics include IHMC’s HexRunner, FastRun-
ner [102], and EllipticalRunner. HexRunner is a rimless wheel with spring-loaded spokes
which generate a spring-mass gait when spinning at a nominal rate. EllipticalRunner has a

Trunk
Velocity Body
Weight
Support
Swing Leg
Pendulum
( Stance Leg
Inverted
Pendulum
Push-Off rt Collision

Figure 6: BlueBiped, a passive-dynamic walking mechanism from Nagoya Institute of Technology.

Strobograph from [81] shows the walking sequence (toe-off torque only applies to actuated
walkers, where BlueBiped is passive and requires an external force to walk continuously).
 2.4 active dynamics through high-transparency actuators 12

Figure 7: IHMC Elliptical Runner gait.

sophisticated mechanism which automatically generates toe trajectories which adapt to

real-world state, not a measured state (Figure 7). EllipticalRunner has only a single motor
which powers two out-of-phase legs. There is significant compliance onboard, both in the
leg-length direction for spring-mass ground forces and in the drive cranks to allow the
legs to become more in-phase under higher load.
Obviously, these heavily under-actuated designs are only effective at a handful of tasks;
their fixed designs are not adaptable to tasks like stairs, stepping stones, or ladders.

2.4 active dynamics through high-transparency actuators

Locomotion is principally concerned with controlling contact forces (Section 4), but high-
ratio gearmotors make force control nearly impossible. A number of smaller robots have
used direct-drive motors and ultralight legs to maximize transparency for locomotion.
Since electric motors are a pure torque source, there is no reason to measure and feed back
on applied force as in an SEA; the commanded force is exactly the contact force, neglecting
the minimal dynamic effects of lightweight links.
Robots such as the MIT Cheetah [154], Jerboa [73], and Minitaur [73] take this approach,
allowing for simplified mechanisms, better contact force transparency, and software-
controlled impedance with almost zero leg inertia. Such designs dramatically reduce
contact impulses and simplify control since they produce almost exactly the contact force
that is commanded. The direct-drive robots can be slightly lighter since they do not require
the steel and bearings of gearmotors.
Force transparency comes at a cost, however, in the form of resistive heating. Electric
motors are most efficient when they are driven at high speeds and low torque, but direct-
 2.5 atrias and cassie 13

drive motors operate at low speeds and high torque. Applying higher torque requires
more current through the motor windings, producing much more heat than mechanical
power. Resistive heating dominates the power requirements of low-reduction robots. For
example, the MIT Cheetah dissipates 75% of its electric power output to resistive heating
[133]. Even so, the MIT Cheetah has a very good total cost of transport at 0.7.
Direct-drive approaches minimize reflected inertia, allowing for easier feed-forward
control, but electric motors are just not ideal for high torque. Perhaps electric motors
will one day be replaced by actuators more similar to biological muscle, which is great
at efficiently producing torque. Muscle is so force-dense that it is generally geared up
(increasing velocity and decreasing force). Direct-drive joints are phenomenally better for
locomotion than geared joints, but we will have to wait for actuator technology to improve
before seeing inertialess, high-torque robotic joints.

2.5 atrias and cassie

I have led the mechanical design of two robots which try and find the right balance between
passive dynamics and software control: the ATRIAS and Cassie bipeds at Oregon State
University (Figures 8 and 9). ATRIAS was designed to have trivial hardware dynamics to
allow for intuitive, heuristic controllers. Cassie was designed to be a more production-ready
robot, intended to be eventually sold by Oregon State’s spinoff company Agility Robotics.
Cassie makes compromises in favor of mechanical simplicity rather than dynamical
simplicity, forcing us to investigate more advanced modeling and control techniques.

2.5.1 ATRIAS

Perhaps the best way to describe the philosophy of ATRIAS is to explain the name: Assume
The Robot Is A Sphere. Introductory physics classes often tell us that everything from
cars to cows can be modeled by simple point masses, allowing us to calculate their gross
motion. Despite being aesthetically absurd, there is a lot of truth to this approach; all
 2.5 atrias and cassie 14

mechanical systems have a center of mass which has a proportionality between its own
acceleration and the forces acting upon the system it represents.
ATRIAS attempts to gather all its mass near the hip joints so that the lightweight
pantograph legs rotate precisely around the robot’s CoM [53, 54]. Series springs connect
the highly geared leg motors to the pantograph, making the overall dynamics of ATRIAS
very close to a simple spring-mass model. Each leg acts as a nonlinear, massless spring
between the ground contact and the "point-mass" robot.
It took a considerable amount of effort and compromise to force all of ATRIAS’s mass
into the hips, but we were sure that it would pay dividends in the form of simple heuristic
controllers [79]. If the robot’s dynamics approximated a simple spring-mass model, then
we ought to be able to control it just like a simple spring-mass model. The idea was
that we could develop heuristics on simulated models with simple dynamics and have
those gaits work on the real robot. However, even when going to extremes to match
reduced-order dynamics, ATRIAS’s dynamics were different enough to invalidate many
of the reduced-order controllers. In the end, ATRIAS required quite a lot of tuning and
human involvement in creating workable controllers [62, 123, 124].

2.5.2 Cassie

After the maintenance nightmare of ATRIAS, we were focused on making Cassie a simpler
robot to operate and maintain. Where ATRIAS had encoders and wires buried underneath
layers of bearings and motors, Cassie’s electronics are all immediately accessible. Each
joint is its own module, allowing for easy replacement of different components. Did a
bad fall destroy the lower leg? Four bolts and a wire bundle is all it takes to remove and
replace.
Cassie has similar series compliance to ATRIAS, with two springs in the lower leg
buffering impacts and storing energy within the leg plane. Interestingly, series compliance
does not have to be directly in series with the motors. On Cassie, the hip is rigidly
actuated, and the knee has a series spring. The knee spring only provides one dimension
 2.5 atrias and cassie 15

Figure 8: ATRIAS at the OSU football field. ATRIAS has only three degrees of freedom per leg.
Abduction raises and lowers the planar leg, and two concentric motors operate the
pantograph in the plane.
 2.5 atrias and cassie 16

Figure 9: The OSU Cassie biped has abduction, long-axis rotation, hip, knee, and toe degrees of
freedom. The knee and ankle have tuned series-compliant elements for efficient and
robust walking.
 2.5 atrias and cassie 17

of compliance, so a second spring is attached farther down the leg on an otherwise passive
joint.
Making mechanically-centered design choices naturally made Cassie’s kinematics and
dynamics more complicated, but not insurmountable. The knee spring joint is not concen-
tric with the knee motor, it is displaced distally to allow it to be connected with a plain
bushing rather than heavier bearings. Also, the heel spring acts as a link in the pantograph
mechanism and deflects under load. These two features make Cassie’s pantograph mech-
anism act more like a six-bar linkage than a four-bar (Figure 10). Controllers can either
neglect these effects, or they can apply loop-closure constraints to the full leg model.
ATRIAS had a serious mechanical flaw that was fixed in Cassie: an antagonistic work
loop during walking and running gaits [3]. ATRIAS’s parallel mechanism required the
two leg motors to exert equal and opposite torque to lift the weight of the robot, while
each motor moves in the same direction to sweep the leg underneath the body (Figure 11).
Opposite torques and equal velocities results in equal-and-opposite power for the two leg
motors, even for nominal spring-mass gaits that require zero net power. This antagonistic
work loop unnecessarily dissipates power in the braking motor, and it increases motor
power requirements.
Cassie uses serial leg motors to avoid antagonistic work for walking and running. Since
the hip strictly swings the leg, it is the only motor that has to move during locomotion,
while the knee just has to support the robot’s weight at zero velocity. As such, for ideal
spring-mass gaits, neither motor is doing any work, and the gait ideally requires no work,
so there is no waste.
Cassie uses actuators optimized for efficiently performing walking gaits [122, 127].
ATRIAS used high-reduction 50:1 transmissions and 170 mm diameter pancake motors,
resulting in substantial reflected inertia. With such high inertia, ATRIAS had to do a lot of
work simply to swing the leg back and forth. Cassie uses an optimal combination of gear
reduction and motor frame size to balance mechanical work output, resistive heating, and
work to accelerate the reflected inertia.
 2.5 atrias and cassie 18

Hip

Knee

Heel

Heel Knee Spring

Spring
Ankle

Toe Actuator

Toe

Figure 10: Cassie’s two series springs are modeled as a planar six-bar linkage (black). The rigid
mechanism moves as a four-bar connecting the hip, knee, ankle, and heel. The plate
spring at the heel deflects with an approximate point of flexure along the spring (point
labeled as "heel spring").
 2.5 atrias and cassie 19

θ̇A τA — 312 W
τB θ̇B
PA 90 J τA θ̇A
PB τB θ̇B
ẋ -92 J

477 W —
— 143 W
Pnet 54 J f · ẋ
-56 J

372 W —

f Swing Stance Swing

Figure 11: Motor power requirements for the ATRIAS robot during a walking task. The parallel
mechanism decomposes leg-length forces into equal and opposite motor torques while
the motors have equal velocity to sweep the leg underneath the torso. This results in
the motors doing equal and opposite work during walking, dramatically reducing the
efficiency of the robot and increasing peak power requirements above what is actually
required by the task. Cassie was designed to correct this issue while maintaining simple
passive dynamics.
3 M O D E L I N G V I RT U A L P R O TO T Y P E S

All design processes require testing and iteration in order to discover and "hone" a good
design. Building designs and testing them in the real world has the advantage of providing
ground truths, but we do not always have the resources to do so. Certain tests require a
real robot (strength and lifetime testing usually sacrifice real components), but often we
can use mathematical models and simulation environments to test a "virtual prototype". It
is often said that simulations are doomed to succeed, so we need to pay close attention
to the models we use to simulate virtual prototypes. The models we use obviously have
to represent reality to a certain extent, so the behavior of the simulated robot reflects the
behavior of the robot as-built.

3.1 software, hardware, and the world

Locomotion exists as a dynamical process within the interactions between three dynamic
subsystems: software algorithms, hardware mechanics, and world dynamics. Each com-
ponent has its own internal state, laws governing its dynamical evolution, and interfaces
through which they can effect change in the others (Figure 12). This work focuses on
the rigid-body mechanics of a robot and touches on the software interface and contact
mechanics in order to make informed, contextual mechanical design decisions. Classically,
hardware and world dynamics are lumped together into the "plant" which interacts with
the controller. However, since mechanical designers have quite a lot of control over the
dynamics of a new robot’s hardware, but they have relatively little control over what
happens in the real world, there is a useful distinction between hardware and world
dynamics.

20
 3.1 software, hardware, and the world 21

Control action
Contact
Controller Hardware World
Proprioception
Perception

Figure 12: Hardware, software, world interaction

hardware Hardware dynamics are governed by the laws of rigid-body mechanics,

receiving internal forcing from the control system and external forcing from the world.
Locomotion hardware typically does not have thrusters (although flying humanoids with
hand and foot thrusters would be spectacular), so the control system has no means of
affecting the robot’s linear and angular momentum directly. Only contact force with the
world can produce changes in total momenta, so the control system must excite the desired
contact forces.

control The control system takes sensor inputs, mixes it with some internal state,
and produces torque commands for the robot’s active joints. The controller has no direct
control over the world, but it can act indirectly through the robot’s hardware. Software can
measure aspects of the world’s state through the robot’s sensors, but that has little bearing
on hardware design so it will not be covered here. This work is concerned with potential
control action on the hardware’s dynamics, not the implementation of specific algorithms.

the world Finally, world dynamics encompass everything happening outside of the
robot: from the contact mechanics between a foot and the ground, to the behavior of a box
that needs to be carried, or even an approaching vehicle that needs to be avoided. This
work is concerned with the effect of external forcing on the robot’s hardware, not any
aspects of mapping, localization, or planning.
 3.2 core mechanical system 22

3.2 core mechanical system

To build a virtual model of a robot, we must encode links and inertia, joint degrees of
freedom, and reflected inertia as a multibody system (Figure 13 a). Multibody dynamics
are mostly concerned with the inertial behavior of a system, since actuator forces, spring
forces, damping, and friction can all be applied as forcing functions to the basic inertial
dynamics. For this work, let us restrict members of the multibody system to be rigid (no
soft bodies). Rigid bodies are far easier to model mathematically [38–41], and they are a
reasonable representation of most real robotic links (excepting soft robots). Rigid body
motion has been throughly studied and we have not only simple governing equations
(Newton-Euler equations of motion), but we have sophisticated algorithms for quickly
simulating arbitrary assemblies [43, 44, 80]. Keep in mind, though, that the mathematical
concept of a rigid body is an approximation of real bodies; there is no such thing as a real
rigid body, and all materials will bend and deform according to continuum mechanics
when sufficiently stressed.
The most basic step in modeling a multibody system is drawing a system boundary:
the imaginary envelope which distinguishes "robot" from "not robot" (Figure 13 b). The
inscribed system has basic quantities associated with it: a center of mass (CoM), a total
mass, and linear and angular momenta. Forces crossing the system boundary (the robot
interacting with the world) modulate the system’s momenta by definition. Forces within
the system boundary (joint torques, for example) do not affect the system’s momenta.
Robot topologies are constructed by connecting rigid bodies together with joints. The
choice and number of joints (rotary, prismatic, multi-DoF) create the robot’s configuration
space, assigning numerical coordinates to poses and shapes the robot can take in the world.
Joints do not always add degrees of freedom to the robot, since a closed loop of connected
bodies has some joints acting as mobilizers and some joints acting as constraints. Robots
with closed-chain topologies have both dependent and independent joints (which joints
are which is a modeling choice), but only the independent joints form the configuration
space. Constructing a model in this way allows us to express its kinematics and dynamics
in minimal coordinates: the positions of the independent joints.
 3.2 core mechanical system 23

I1a1 v1 I1v1
fy
fx Aq + Aq

I3a3 v3 I3v3
I2a2 v2 I2v2

(a) f (b) f

Figure 13: (a) shows a multibody system with an externally-applied force, link accelerations, and
joint forces. By drawing a system boundary, (b) we can capture the relationship between
external forces and gross system motion at/around the centroid. See [104, 105].

Computing kinematic quantities for a chosen topology is a matter of mapping position,

velocity, acceleration, and force vectors between joint coordinates (q, q̇, q̈, τ ∈ Rn ) and
world coordinates (x, ẋ, ẍ, f ∈ Rm ). For any joint velocity of a manipulator, there is an
associated end-effector velocity ẋ = Jq̇. For any load applied to the end-effector, there is
an associated set of joint torques τ = J⊤ f. Such mappings are usually uniquely invertible
if the dimension of the joint and task spaces are the same (n = m), but they are often
underdetermined for articulated mechanisms. The Jacobian matrix J(q) can be interpreted
as a row of task-space basis vectors for the joint variables qi . These basis vectors ei are the
partial derivative of the function mapping joint space to task space x(q).

∂x(q)
ẋ(q, q̇) = q̇ = J(q)q̇ task-space velocity (1)
∂q
X ∂x
= q̇i partial derivatives (2)
∂qi
X
= ei q̇i joint basis vectors (3)
 3.2 core mechanical system 24

These basis vectors are equivalently used to map task-space forces into joint space

τ⊤ q̇ = f⊤ ẋ conservation of power (4)

τ⊤ q̇ = f⊤ Jq̇ replace task velocity (5)

τ⊤ = f⊤ J must be valid for all q̇ (6)

τi = ei · f joint torque as a scalar product (7)

where the torque applied to each joint in the task space is equivalent to the scalar product
between that joint’s basis vector and the applied task force. The basis-vector interpretation
of the Jacobian matrix is used in Section 5 to derive visual representations of locomotion
metrics.
This work models a robot’s mechanical system using standard joint-space dynamic
equations,

M(q)q̈ + h(q, q̇) = Bu + J(q)⊤ f, (8)

where M is the joint-space inertia matrix relating joint accelerations q̈ to joint forces τ. h is
the vector of velocity-product and gravitational forces ("bias" forces). B is a selector matrix
mapping control action u into joint space. Unless otherwise stated, joint coordinates will
always be those of the articulated body tree and not the input-side torque and velocity of
actuators. So, actuator output torques in u will directly affect joint-space dynamics, and
the B matrix will only have unit and zero entries.
Dynamics equations are written symbolically, but that is not to say that the terms
must be symbolic. Lagrangian [15] and Hamiltonian mechanics work well for simple
systems, generating symbolic equations of motion in minimal coordinates either "by hand"
or through a computer-algebra system (e.g. MATLAB Symbolic Toolbox or Python’s
SymPy). Alternatively, systems with more complex dynamics and constraints are amenable
to numerical dynamics (see Featherstone’s book on dynamics algorithms [42]). Many
simulation packages exist for calculating kinematic and dynamical properties such as
the MATLAB Robotic Systems Toolkit, the Rigid Body Dynamics Library (RBDL [44]),
 3.2 core mechanical system 25

MuJoCo [145], and so on. There are also many special-purpose geometric, symplectic, and
Lie-group integrators for generating accurate system trajectories [87, 110].
It is easiest to calculate the terms in the joint-space dynamics equation for a tree topology,
but not all robots are open chains. To include constraints, we start with an open-chain
system and include Lagrange Multipliers λ as constraint forces as well as the restriction
that the constraint’s position, velocity, and acceleration must be zero,

Mq̈ + h = Bu + J⊤ f + G⊤ λ dynamics with constraint forces (9)

g(q) = 0 constraint position (10)

ġ = Gq̇ = 0 constraint velocity (11)

g̈ = Gq̈ + Ġq̇ = 0 constraint acceleration (12)

where G is the Jacobian of the holonomic constraint function g(q). q and q̇ must be
initialized so as to satisfy Equations 10 and 11. Terms in this equation can be found
symbolically or by using the Composite Rigid Body Algorithm to find M and the Recursive
Newton-Euler Algorithm to find h and J⊤ f. Solving the constrained system can be done
by formulating a system of linear equations in q̈ and λ,

    
M −G⊤ ⊤
 q̈  Bu − h + J f 
   =  , (13)
G 0 λ −Ġq̇ − αg − βġ

where α and β are Baumgarte stabilization parameters which drive g(q) to zero when the
robot’s configuration does not completely satisfy the constraints [45].
Alternatively, we can project the open-chain dynamics onto the constrained dynamics to
yield a system in minimal coordinates. The projection approach yields a single dynamic
equation that is constraint-agnostic

Mγ q̈ind + hγ = Bγ u + J⊤
γ f. (14)
 3.2 core mechanical system 26

We can arrive at this equation by finding a matrix γ such that Gγ = 0. One way of
generating γ is to partition G into columns associated with the independent joints qind
and the dependent joints qdep , then writing q̇ in terms of q̇ind alone

   
 q̇ind   q̇ind 
 
Gq̇ = Gind Gdep  =0 =⇒ q̇ =   = γq̇ind , (15)
q̇dep q̇dep
 
I
where γ =  . (16)
 
−1
−Gdep Gind

So long as the constraints are not redundant, there will be as many dependent joints as
constraint equations, so Gdep will be square and invertible. Another way of finding γ is to
write a function mapping independent coordinates into full coordinates as in q = y(qind ).
γ is then simply ∂y/∂qind since q̇ = ẏ = γq̇ind .
We can use γ⊤ to project the unconstrained dynamics into the constrained dynamics
(removing the Lagrange multipliers) as in

γ⊤ Mq̈ + γ⊤ h = γ⊤ Bu + γ⊤ J⊤ f + γ⊤ G⊤ λ (17)

= γ⊤ Bu + γ⊤ J⊤ f + 0. (18)

where Gγ = γ⊤ G⊤ = 0, which is proven simply as

 
I
 
Gγ = Gind Gdep   = Gind − Gind = 0. (19)
 
−G−1
dep Gind

We can cast q̈ in terms of q̈ind by seeing that

 
 q̈ind 
 
g̈ = Gq̈ + Ġq̇ = 0 =⇒ Gind Gdep   + Ġq̇ = 0 =⇒ (20)
q̈dep
     
 q̈ind   I  0 
q̈ =  =  q̈ind +   Ġq̇ = γq̈ind + γ̇q̇ind . (21)

−1 −1
q̈dep −Gdep Gind −Gdep
 3.3 control interface 27

Ġq̇ is a vector which is self-explanatory to calculate symbolically, but it may be more

intimidating when using dynamics algorithms. The knee-jerk reaction may be to use
finite-differencing on G, but evaluating the constraint acceleration function g̈(q, q̇, q̈) with
q̈ = 0 returns the vector Ġq̇.
Combining Equations 18 and 21 produces

γ⊤ Mγq̈ind + γ⊤ Mγ̇q̇ind + γ⊤ h = γ⊤ Bu + γ⊤ J⊤ f, (22)

giving us the terms for the projected dynamic equation

Mγ = γ⊤ Mγ, constrained inertia (23)

hγ = γ⊤ (Mγ̇q̇ind + h) , constrained bias force (24)

Bγ = γ⊤ B, constrained actuator matrix (25)

Jγ = Jγ, constrained Jacobian (26)

Mγ q̈ind + hγ = Bγ u + J⊤
γ f. (27)

The rest of this text discards the γ subscript and treats both constrained and unconstrained
dynamics with the same equation, simply requiring that if a design contains constraints
that its dynamics be projected into the constrained subspace.

3.3 control interface

It is the control system’s job to excite motion and generate locomotion gaits, moving
the robot through the world. Hardware dynamics and the control system have been
deliberately partitioned to allow for passive mechanical components to be included in the
"control system", since their behavior is very much like a simple, fixed controller acting
on the joints. This reduces the complexity of the hardware’s dynamical equations and
reinforces the use of "mechanical intelligence" [21].
The control system interacts with the hardware by generating internal forces using active
mechanical actuators and passive components like springs and dampers. This work focuses
 3.3 control interface 28

on electrical actuators (brushless DC motors with gearheads) and joint-level passive compli-
ance. Including actuators in a mechanical system introduces acceleration-proportional force
at the joint (reflected inertia), and including springs introduces deflection-proportional
force. Passive compliance is nearly lossless, but generating torque with actuators carries
with it an energy-loss model.

3.3.1 Actuators

Actuators give the control system a channel through which to apply arbitrary forces
(within saturation limits) to the mechanical system. Adding actuators comes at a cost,
however: not only do they increase the weight of the system, but they add reflected inertia
and require power to exert forces. They are required for machines that do work and adapt
to their environment, but in some cases passive components are more appropriate.
Electric actuators have a relatively simple loss model, requiring electrical power to be
converted into mechanical work (that which is necessary to accelerate the reflected and
link inertia, resist viscous friction, and do work on the world) and also to generate holding
torque.

v = kb θ̇ + iR winding voltage (28)

i = τb /kb winding current (29)

τb = τ + Iθ̈ + cθ̇ magnetic torque (30)

Pelec = iv = τb θ̇ + τ2b /k2W electric power (31)

= τθ̇ + Iθ̈θ̇ + (τ + Iθ̈ + cθ̇)2 /k2W (32)

where kb is the winding constant (identically the torque and back-EMF constants) of the
motor and kW is the motor constant relating torque to winding losses in resistive heating.
 3.3 control interface 29

τ is the actuator’s output torque, θ̇ and θ̈ are the actuator’s speed and acceleration, and I
is the inertia of the actuator. I and kW for a motor/gearhead pair are calculated simply,

I = Ils + G2 Ihs (33)

c = cls + G2 chs (34)

kW = Gkm (35)

where G is the gearhead reduction factor, Ils is the inertia of the low-speed side of the
actuator (output, typically), Ihs is the inertia of the high-speed side (motor rotor), c is
viscous friction, and km is the motor constant.
Geared motors introduce reflected inertia at the actuated joint, increasing that DoF’s
resistance to acceleration. This is not a spatial inertia that resists body acceleration, but
strictly acts on the relative acceleration between two links connected by an actuator.
Adding this reflected inertia to the mechanical model is as simple as appending an Ii to
the diagonal element of M indexed by the joint to which the actuator i is attached (an
actuator on the second joint of a 2-link serial manipulator would append M2,2 ).
Actuators are treated as torque sources at chosen joints in an articulated rigid-body
system. Since we have discussed loop-closure constraints, there is no reason for actuators to
"bridge" multiple joints, since the constrained Bγ matrix will account for any multi-articular
effort.

3.3.2 Passive compliance

Springs are best thought of as fixed proportional controllers on a robot’s joints. The
advantage of this over active compliance is that passive compliance does not have any
energetic costs and no reflected inertia or other actuator dynamics (backlash, friction, etc).
So, if part of the software control system would only ever generate a fixed force/deflection
profile at a joint, it is better to use a physical spring at that joint. Springs introduce a
predefined forcing function on the joints as a function of joint deflection. They can be
 3.3 control interface 30

(a) (b)

Figure 14: (a) uses nonlinear cantilever-beam deflection and a link to get approximately linear
torque/deflection response around the compliant joint. (b) approximates the large-
deflection beam behavior using one or more joints with linear spring functions.

in series or parallel with motors [46], and they can be mono- or multi-articular through
loop-closure constraints.
The Cassie and ATRIAS robots use cantilevered fiberglass plate springs, posing an
interesting modeling problem. Both ATRIAS and Cassie have spring linkages which make
the cantilevered spring behave like a proportional spring around the compliant joint
(Figure 14 a), but Cassie also uses a cantilever spring as a link with no fixed flexure joint
(Figure 14 b). This may seem to introduce modeling difficulty, but there is a trick: such
cantilever springs can be approximated as a series of rigid links with proportional springs
at the joints [142].
We can identify the spring rate of compliant joints in the real robot by using a fitting
process. One option is to identify the spring rate of each joint individually, and another
is to identify all compliant joints at once by correlating force/deflection response of the
end effector in task space. When using the multi-link approximation of a cantilever beam,
the distances between flexure points are also open variables that need to be discovered
based on test data. The test samples joint positions and contact forces (using a forceplate or
load cell) for a range of feed-forward motor torques. The fitting process tries to find linear
spring rates and flexure positions in order to minimize the error between the joint torque
induced by the measured contact force and the joint torque modeled by the deflection of
joint compliance.
 3.4 world interaction 31

3.4 world interaction

Since locomotors do not have thrusters, contact is the only source of controlled motion
through the world. Unfortunately, we generally have no control over ground composition
in the real world, and contact in general presents substantial modeling difficulty. However,
we can still make useful mechanical design decisions by looking at the contact interface
and leaving the contact model as a black box, permitting variable ground composition and
friction. For example, we can devise control systems that are insensitive to touchdown
timing and ground composition [62, 124].
Contact can be modeled as a discrete impact event followed by continuous unilateral
forcing [138]. Before contact, the robot’s end effector and the world have a relative velocity
and a non-zero separation distance. At the moment of impact (separation reaches zero),
each body experiences an impulse to bring their relative velocity to zero. The duration of
contact (compression/restitution) occurs in continuous time, applying non-zero external
force to the robot according to the contact model (rigid, soft [144], deformable, etc). Contact
terminates at the moment when applied force crosses zero† .
Contact force on a body is a 6-element spatial force (force and moment) [40, 41] which
typically does not allow a Center of Pressure outside the contact area. Some sort of
adhesion would be necessary between the end effector and the world to allow the CoP to
be outside the contact area‡ . The spatial contact force can be constructed using the sum of
linear contact forces at different points on the contacting body. This is useful for robots
like Cassie which have line-contact feet [1], since the zero-roll-torque constraint becomes
implicit due to only two contact points on the foot.
The real robot must be built and controlled to permit uncertain contact models and
touchdown timings (sand, mud, loose rubble, ice, etc). The only thing we generally know

† In general, the correct way to handle the contact state is to establish contact based on distance and to release
based on force [70]. This prevents the accidental inclusion of adhesion effects where a velocity-dependent force
balances a compression-dependent force before the compression distance becomes zero.
‡ The Center of Pressure is really the intersection of the "line" aspect of a spatial contact force with a "support
polygon". Support area makes less sense when hands and feet are used on terrain with changing or uncertain
impedance. It makes even less sense when hands are allowed to grip the environment, presenting bilateral
forcing. Dynamic locomotion is only concerned with the net contact force, not its intersection with an arbitrary
polygon.
 3.4 world interaction 32

beforehand is that the contact state will "flow" in the direction of commanded contact force.
That is, the robot’s limbs will extend when the commanded force cannot be matched by the
current state of contact. Rigid contact makes this instantaneous, supplying the restoring
force immediately, but granular media and slippery surfaces will have some delay (or may
not be able to support the commanded contact force due to friction or other limitations).
Shaping this behavior through mechanical design allows us to use control schemes which
"fail gracefully" when they encounter unexpected circumstances. In practice, friction cones
are used more to keep contact forces within some safety margin, not to actually model
contact friction.
Terrestrial systems are subject to gravity, but those systems only feel gravity when they
resist it. It is the contact force, then, which couples the effect of gravity to acceleration
of the robot’s joints† . This fact presents many interesting opportunities for control, such
as "gravity compensation" which accounts not just for a vertical 9.81 m/s acceleration,
but also virtual forces resulting from non-inertial reference frames (e.g. traveling along a
winding road in the back of a truck).

3.4.1 Touchdown event

For any given mechanism, we know the relationship between task impulse and task
velocity delta as

H = J⊤ Λ joint-space impulse (36)

∆q̇ = M−1 H = M−1 J⊤ Λ joint-space velocity delta (37)

∆ẋ = J∆q̇ = JM−1 J⊤ Λ task-space velocity delta (38)

where (JM−1 J⊤ )−1 (which can also be calculated using dynamics algorithms [155]) is the
Operational Space Inertia Matrix (OSIM) [14] and Λ is the task-space impulse. This matrix
(and the impulse-velocity relationship it represents) can be visualized as an ellipsoid in

† For example, arm swing is driven by ground-reaction force on the feet, and forceful arm motions spike the
ground-reaction force.
 3.4 world interaction 33

task space, where major diameters represent the "heavy" directions and minor diameters
represent the "light" directions.

3.4.2 Continuous forcing

Rigid ground implies a motion constraint that allows us to solve directly for contact force

x(q) world position of contact (39)

ẋ = Jq̇ = 0 contact velocity constraint (40)

ẍ = Jq̈ + J̇q̇ = 0 contact acceleration constraint (41)

J⊤ f = Mq̈ + h − Bu equation of motion (42)

M−1 J⊤ f = q̈ + M−1 (h − Bu) invert mass (43)

JM−1 J⊤ f = Jq̈ + JM−1 (h − Bu) left-multiply Jacobian (44)

JM−1 J⊤ f = −J̇q̇ + JM−1 (h − Bu) Jq̈ = −J̇q̇ (45)

 −1 
f = JM−1 J⊤ rigid contact force (46)

JM−1 (h − Bu) − J̇q̇

Alternatively, we can use an augmented inertia matrix to simultaneously solve for joint
acceleration and contact force
    
M −J⊤ q̈  Bu − h
  =  , (47)


J 0 f −J̇q̇ − α(x − x0 ) − βẋ

where the Baumgarte stabilization parameters α and β are necessary to stabilize contact
distance and velocity during simulation.
4 D E F I N I N G T H E L O C O M O T I O N TA S K

Our design process needs a precise definition of what candidate designs are supposed to
do. With a clear definition of purpose, we can design the hardware to ensure the feasibility
of a robot, increase efficiency, and reduce the stability burden placed on the control system.
Additionally, specifying a task allows us to pick appropriate objectives for guiding our
search through the design space.

4.1 abstract locomotion

In general, the purpose of a locomotor is to transport itself from one location to another
over potentially rough or hazardous terrain. So, let us choose to initially define locomotion
as the self-transport of a system of connected bodies. Self-transport means that the system
should be able to move through space without the assistance of external devices. Everything
not connected to the system is considered as "the world".
In order to transport its own mass through space, the locomotor system has to adjust its
linear momentum. Simply, speeding up requires an impulse in the desired direction and
slowing down requires the opposite impulse. Additionally, we can say that locomotion is
associated with zero average angular momentum, since any undulation is periodic and not
continuous (barring acrobatics). Only forces acting across the system boundary can affect
the system’s momentum, since forces within the system simply exchange momentum
between components and will not affect the total.
For terrestrial locomotion, the system will have no propellers or thrusters, so the only
source of controlled impulse is the force of contact with the world. This allows us to say
that terrestrial locomotion is ultimately the act of exciting contact forces and impulses.

34
 4.1 abstract locomotion 35

This statement is made with no loss of generality relative to the original definition of mass
self-transport.
Such a definition may appear to neglect attitude adjustment during free motion (no
contact), but it is often necessary to reorient and reconfigure the robot in order to apply
desired contact forces in the future. Reorientation takes advantage of the conservation of
angular momentum to reorient the system rotationally [88]. Total angular momentum still
can only be adjusted through contact force.
As a side note, controlling contact force is different than merely being able to resist
contact forces; a chair’s legs are in contact with the ground, and the combined impulse of
their contact perfectly balances the impulse of gravity, but that does not make it a good
locomotor. The difference lies in whether the commanded actuator torque resists applied
loads or if it is the acceleration or compression/tension of passive components; torque due
to magnetic flux in the air gap of an electric motor is controllable whereas torque due to
other effects is not. Robots can gain efficiency by using passive components where possible
and using active components only where control and adaptability are necessary. Robots
can also exploit singularities (which are typically avoided because they limit the robot’s
control authority) to resist contact forces rather than spending energy on motor effort.

4.1.1 Abstract control structure

In order to control its motion, a system must be able to excite contact forces (the system’s
only source of impulse). By definition of the mass-center of a system, both controlling
momentum and controlling contact force are equivalent to controlling the acceleration (or
change in velocity) of the system’s center of mass. It follows that, in general, the control
system of a locomotor is interested in commanding both linear and angular centroidal
(center of mass) accelerations.
Since locomotion is ultimately the combination of exciting contact forces and recon-
figuring the robot’s pose, "locomotion-level" commands can be sent to an interface layer
which abstracts from the specific hardware design. A fair amount of research has been
 4.2 spring-mass locomotion 36

done on such an interface as "whole-body control" [71, 105], "task-space control" [28, 89],
or "operational-space control" [76]. Such control interfaces take acceleration and force com-
mands for the centroid or other points on the robot and deduce the joint torques necessary
to produce those accelerations given ground contact, friction cones, and other constraints.
Whole-body locomotion control is only concerned with the CoM motion task, leaving room
in the controlled dynamics for other tasks (e.g. manipulation, acrobatics, body language
communication, etc). One could imagine this interface as a more sophisticated version of
Jacobian-Transpose control [24].
In operation, the acceleration control interface abstracts higher-level controllers from
the kinematics/dynamics of the hardware. Effects of hardware changes are bounded to
this interface layer and do not require changes to the higher-level controllers acting on this
layer. For example, the trajectory generator and path planner can operate on the reduced
point-mass dynamics of the robot’s CoM without needing to know anything about the
robot’s internal dynamics [33].

4.2 spring-mass locomotion

Abstract locomotion places few limits on the dynamics of a locomotor, so we can use
a specific model to usefully constrain the system’s dynamics. At this point, locomotion
only involves the motion of the CoM and the net contact force (i.e. CoP). Arguably, the
simplest dynamic constraint is that the contact force is controlled to increase as the CoM
gets closer to a desired contact point within reach of the system. This causes the system
to "bounce" off of any surfaces with which it comes into contact (imagine potential-field
obstacle avoidance [75]). The Spring-Loaded Inverted Pendulum (SLIP, Figure 15) [20]
codifies these "spring-mass" dynamics.
There is substantial evidence that spring-mass locomotion has benefits in terms of
energy economy [4, 125], self-stability [49], and impact robustness. The SLIP and other
reduced-order models give insight into how robots may exploit natural dynamics for
 4.2 spring-mass locomotion 37

Figure 15: Spring-loaded inverted pendulum (SLIP) model of a human running gait. This minimal
spring-mass model has been shown to produce both walking and running gaits for
humans, as well as realistically model a number of terrestrial animals. Adapted from
[50, 131].

locomotion. By designing a robot to dynamically resemble a SLIP, it can receive the same
benefits and use similar control theories as the SLIP [63].

4.2.1 Reduced-order model

The SLIP can be seen as a simple, lumped-parameter representation of animal limbs. The
SLIP comprises a single point-mass body and a massless linear spring for each leg. Being
composed of only a point mass and springs, the SLIP is conservative and has no rotational
inertia or angular momentum, and the only control signal for the unactuated SLIP is the
leg angle at touchdown. This is not the only reduced-order model for locomotion [34, 59,
116, 130, 135], but it is arguably the most applicable [20, 96]. Additionally, there are quite a
few locomotion control ideas stemming from simple leg angle control of the SLIP [11, 13,
37, 78, 79, 91, 112, 134, 148, 149, 157, 160]. SLIP dynamics require numerical integration
(e.g. [128]), but there also exist closed-form approximations [131, 136].
Amazingly, the simple, unactuated, bipedal SLIP reproduces the CoM trajectory and
ground-reaction forces of human walking and running [49, 50]. However, even though it
is analogous to the CoM and series compliance in animals, the SLIP says nothing about
how animals use physical compliance. We know that animals do indeed incorporate and
exploit compliance [7, 8, 10], and it is possible that animals actively track a reduced-order
 4.2 spring-mass locomotion 38

model [19, 47], but we do not know exactly why. So, we need to remember that robots are
not animals, and biomimicry design arguments are no better than gut instinct.
The SLIP has been shown to have a fragment of self-stability when being controlled with
feed-forward touchdown leg angle [48]. That is, the SLIP can reject small disturbances
passively without feedback control and converge to the original equilibrium gait. Self-
stability suggests that real robots which instantiate SLIP dynamics will have a reduced
control burden due to the passive rejection of disturbances.

4.2.2 Application in design

Mechanical designs can physically approximate SLIP dynamics and receive a number
of the SLIP’s desirable properties. Of course, a real robot must have distributed mass,
so no design will ever perfectly match the SLIP, but much of the intuition still applies.
Many implementations simply copy the SLIP’s "pogo-stick" morphology, thereby trivially
approximating SLIP dynamics [5, 12]. However, articulated mechanisms have benefits over
linear slides [109], so other robots have adopted non-obvious morphologies with the same
dynamic properties [63, 109]. Approximating SLIP is another great example of designing a
machine with dynamics and control in mind [2].
Although compliance for spring-mass gaits could be implemented as a proportional
controller in software [35, 73, 154], it is better to utilize physical compliance due to the
inefficiencies and reflected inertia of motors. Utilizing physical springs as inertialess,
efficient proportional controllers carries a number of benefits (and also a few detriments).

+ Resonant locomotion following spring-mass gaits can recycle gait energy using passive
springs as a high-efficiency energy store [4]. These gaits have a known positive and
negative work requirement over time, and tuned series compliance can do this work
with very little loss as opposed to electric motors which do not easily regenerate
energy and require power simply to resist forces.

+ Passive springs decouple actuator and body inertia from ground impact, reducing force
spikes from unexpected contact. Softening impacts can increase the lifetime of gears,
 4.2 spring-mass locomotion 39

electronics, and other components, and it also reduces the disturbance (discrete
velocity change) associated with large impulses.

+ Leg compliance makes the ground contact model matter less because the known leg
model dominates. This reduces perception and planning requirements, making the
controller simpler.

+ Spring-mass gaits can reduce simultaneous positive and negative work if at least one
actuator operates the leg angle exclusively [3, 8, 81].

+ Electric motors cannot do work at max torque, but series springs can. The motor only
needs to hold the neutral position of the spring, and the spring alone will store and
release energy [55].

+/- Parallel springs help energy efficiency because the motors do not have to exert
themselves, but they limit possible tasks because the motors must fight to maintain
poses away from the springs’ neutral positions. This is a specialization/functionality
trade-off.

- Too much compliance between various inertias (reflected, heavy links, etc) can lead to
several vibration modes if not designed correctly [83], making the robot difficult or
impossible to control.

- Since the goal of locomotion is to excite contact forces, the combination of compliance
and reflected inertia hurts controller bandwidth. Acceleration is now a function of
position, and thus not directly controllable by the actuators. This can lead to ringing
in position and force commands. However, reducing reflected inertia brings the
series-elastic actuators [114] back to being simple torque sources, just like a plain
motor.

Encoding these properties into a limb design is a matter of finding the right "program"
for intrinsic compliance (deflection vs force at the end-effector) and the right inertia
distribution. Even when the right theoretical spring rates are found, they may need to be
nonlinear or otherwise difficult to translate the spring function into physical components.
 4.3 through-contact trajectory optimization 40

4.3 through-contact trajectory optimization

Walking, running, or other tasks can be defined by objectives and constraints for a dy-
namical system. Obviously, simulating the full robot with final controller gives the best
indication of the operating point of a design, but we are often forced to create initial
machines that will work amicably with unspecified, future controllers. If we have access
to a through-contact trajectory optimizer [98, 99, 113, 121] as a stand-in for a controller,
then we can generate candidate trajectories on a design-by-design basis to maximize goals.
However, gait optimization is quite the subject on its own, so this work only discusses it
for completeness.
Given a sufficiently realistic simulation, a computer optimization process could find
very concrete designs that exploit kinematic singularities, have autonomous swing-phase
dynamics, use toe-off forces for leg recirculation, and other passive-dynamic features for
very specific tasks. There has been substantial work in this area, especially in locomotion
[36, 57, 60, 61, 153].
Optimizers find the theoretically-optimal path given perfect information. In reality, the
trajectories will surely be worse, requiring feedback control to stabilize [68]. However, we
are after the peak theoretical performance of a hardware design, so we are not as worried
about tracking.
In the world of optimal control, it is either possible to meet constraints or it is not
possible (feasibility), and it is up to the objective function to determine which feasible
trajectory is best. The objective must provide enough regularity to find feasible solutions,
but it must also lead to the desired solutions. Minimizing the integral of squared torque is a
common objective, giving smooth trajectories fairly regularly (minimizing nondimensional
work is also common [121, 141]). Squared-torque happens to measure the deviation from
passive-dynamic paths (trajectory "error" from those which require no control effort), and
is also proportional to resistive-heating losses in BLDC motors. So, this metric minimizes
a large component of the robot’s power requirements in addition to providing smooth,
least-action trajectories.
 5 PERFORMANCE METRICS

Objectives guide our search for better locomotor designs by indicating how well we expect
possible designs to function relative to each other. Even without specifically setting these
down, designers intuitively use objectives and metrics to judge designs. For example, we
are commonly interested in reducing weight and component cost while increasing load
capacity (making the robot lighter, stronger, and cheaper). Meticulousness encourages us
to lay out all the factors at play in our design program, so that is what we will do.

5.0.1 Properties of good performance metrics

In general, a metric can be any qualitative or quantitative evaluation of a real property

of a robot which provides the ability to rank candidate designs. A metric is typically a
well-ordered quantity (like Integers or Real numbers), but a they can be any formal set of
weakly-ordered objects‡ . We require the set-theory definition of metrics when we begin to
combine disparate metrics and identify over-all better designs in Section 6.
Metrics should indicate tangible, real-world performance, so there must be a mapping
from real designs into the metric space. Without the connection to a specific, real property
of the system, metrics just create noise in the design process. So, it is important to calibrate
metrics against real properties. Additionally, if the chosen metrics for a project lead to
sub-par results, then the project metrics needs to be re-evaluated.
Metrics must be normalized in order to fairly compare robots of different weights,
sizes, and speeds. For example, Jacobian-based performance metrics are susceptible to
unit inhomogeneity when both linear and rotary actuators are used simultaneously [90].
Without homogenization, robots of different types are not comparable. Similarly, it is

‡ Weak ordering means each element of a set is either less than, equal to, or greater than any other element of
the set.

41
 performance metrics 42

much more impressive for a tiny, 10 cm-long hexapod to travel at 1 m/s than it is for
a 1.5 m cheetah, so simply measuring top speed is not a good metric. In most cases, a
properly-normalized metric has no units.
If it is possible, bounded metrics provide upper and lower limits which give a sense of
where a particular design falls in the spectrum of possibilities. Of course, all numbers are
bounded by infinities, but that is the default. If we can bound metrics between meaningful
upper and lower limits (or at least provide reference values), we can gain a better insight
into the possible designs.
Metrics should also be smooth, if possible. Smoothness implies that a subtle change in
the input design produces a subtle change in the metric. Discontinuities in the mapping
between designs and their metric value can lead to confusion and make it difficult to hone
into an optimal design. For example, mechanical linkages have singularities which have a
large impact on manipulability metrics. As a general rule, metrics should be functional up
to and through singularities.

5.0.2 Locomotion metric categories

When creating a new hardware design for locomotion, there are four categories of objectives
that have been described by various groups and are generally agreed upon:

1. The energetic efficiency of a design in performing its tasks, measured by the drain
on its power supply. Energy efficiency is critical for mobile robots, because they must
carry their own power source.

2. The control authority of a design, measuring how "effectively" the control system
can affect change in its control target. Classical manipulability metrics are placed
here and assume force, velocity, or acceleration control targets of the end-effector.

3. The robustness of a design to control errors and external disturbances. Slips, impacts,
and uncertainty cannot be allowed to threaten the stability of the gait. Some designs
are less robust than others, so we need to quantify that in our design process.
 5.1 efficiency metrics 43

4. The feasibility of a design, making sure it is physically capable of walking, running,

jumping, and lifting according to the task description. These objectives are a subset
of the design constraints which are written as penalty functions.

These categories are fairly abstract (perhaps except for efficiency), but we can dive into
each one to find concrete design metrics which promote the higher-level goal.

5.1 efficiency metrics

Mobile robots must strive to increase their efficiency, since they must carry their own power
supply. Higher power requirements lead to larger motors and power sources, increasing
the weight of the robot and leading to even higher power requirements; it is a vicious
cycle. The trouble is, a machine alone does not have power requirements, only does the
combination of a machine and a motion/torque trajectory. As such, we cannot optimize
the efficiency of a machine without specifying how it will be moving and interacting with
the world. The task specification can be as complex as a distribution of speed/torque
trajectories, or it can simply be a pair of force/velocity vectors at the end effector.
Given that power generally cannot be regenerated once spent, the loss model for an
electric actuator is the integral of positive electrical power [141],

Z Z
τθ̇ + Iθ̈θ̇ + + cθ̇2 + (τ + Iθ̈ + cθ̇)2 /k2W (48)
 
P+ =

We want to minimize mechanical power expenditure, and limit costs related to generating
torque/motion. Better regenerative braking will certainly increase the efficiency of gaits
[100, 158], but it is always better to keep power requirements low than rely on regeneration.
Even if power regeneration was 100% efficient, motors still have to be physically larger to
accommodate higher power requirements.
While total positive power output is useful for comparing robots with some common-
ality, the Total Cost of Transport (TCoT) is a unitless measure of power and work that
can compare wholly different robots, species, and even vehicles (although there is a
slight downward trend when comparing large differences in mass [81]). Measuring non-
 5.1 efficiency metrics 44

dimensional work, TCoT is the ratio of work done by a robot’s power source to the product
of the robot’s weight and distance traveled. Similarly, non-dimensional power is simply
the ratio of power to the product of weight and velocity. TCoT is lower-bounded by zero,
implying no cost to locomote. For reference, humans have a TCoT of 0.2, the ATRIAS robot
has a TCoT of 1.0, ASIMO has approximately 3.2, Oregon State’s Cassie and MIT Cheetah
have 0.7, and the Collins walking robot matches humans at 0.2 [32].
Electric motors generally have to be geared down in order to produce enough torque to
drive human-scale machines. Speed reduction at first approximation linearly exchanges
speed for torque while keeping mechanical power constant. However, gear reduction also
has an effect on resistive heating losses and viscous losses. In the sense of electrical power,
the gearing nonlinearly exchanges mechanical work of rotor acceleration and viscous
losses with torque-related electrical losses. Because this exchange is nonlinear and convex,
there is a "sweet spot" when selecting a motor/gearhead pair for a particular application
[22, 122, 127].

5.1.1 Antagonistic work

A robot’s kinematics play a large role in determining power requirements for a machine,
because they determine how the net task power is distributed among the individual motors
[3]. Because locomotion is the self-transport of an inertial system, it has a defined net power
requirement for modulating external kinetic energy (due to the velocity of the system’s
mass-center) and potential energy (raising and lowering the mass-center). And, because
the robot is composed of individual joints which all have their own speed and torque
requirements as defined by the robot’s kinematics, this net power can be produced in a
combination of different positive and negative components. In fact, the chosen kinematics
of a robot as they relate to a task can cause some motors to do substantially more work
than the task requires, and that excess work is offset by braking in other actuators (as noted
since the inception of robotic limb design [58, 108, 139, 151]). The idea of antagonistic
work offers a more complete view than simply "minimizing positive work". It describes
 5.1 efficiency metrics 45

Figure 16: Hardware designs must be "aligned" with a task in order to perform optimally. Above is
a simple demonstration where two identical x-y tables must resist an applied load while
moving at constant velocity in the plane. The only difference between configuration (a)
and configuration (b) is a 45 degree rotation of the table relative to the force and velocity
vectors. As a result, table (a) must work harder to do the same task.

how much work is being done that does not benefit the task, whether the task requires
positive or negative work. Figure 16 gives a simple example.
Measuring antagonistic work is a matter of comparing the total power Σ|Pi | being gen-
erated and dissipated within a mechanism to the net power |ΣPi | required by the task.
Antagonistic work is then half the difference between total and normed net power (the
factor of two is there because antagonistic work is by-definition a pair of equal flows into
and out of a system)

1
Pant = (Σ|Pi | − |ΣPi |). (49)
2

Since antagonistic work does not benefit the task, we need to design mechanism+task
pairs which minimize the amount of antagonistic work. Minimizing antagonistic work is
equivalent to minimizing mechanical cost of transport (MCoT) under a zero-regeneration
motor power model.
 5.1 efficiency metrics 46

Minimizing the amount of antagonistic work in a mechanism+task pair means finding a

set of kinematics which require same-sign work from all actuators. Additionally, balancing
power among all actuators reduces their individual sizes and makes gait variations less
likely to require more power than necessary [3]. Balancing power among all actuators
increases the power quality of a machine+motion pair. Power quality is a smoothed
version of antagonistic work which promotes power sharing between actuators

Q = (ΣPi )2 − Σ(Pi 2 ). (50)

which, as you may notice, is proportional to the variance in P around the mean ΣPi /n,
where n is the number of actuators. The less varied the actuator powers, the more likely
they are to be the same sign, with the limiting case of zero variance distributing the task
power evenly among all the actuators.
For a system with two actuators, power quality reduces to a saddle centered at zero
(Figure 17). By maximizing power quality for a machine+task pair, we minimize antagonis-
tic work as well as the total power that must be generated/dissipated by the actuators,
thereby reducing the required size of the actuators and power supply [3]. Figure 18 shows
how three different leg designs compare in following a spring-mass running gait.
Power quality optimization carries with it a helpful guarantee: finding power-optimal
kinematics guarantees that each actuator will have the smallest power requirements,
allowing joint-level actuator optimization on a joint-by-joint basis. Maximizing power
quality is a kinematic operation which can be done early in the design process, before
actuator internals are even specified. Doing so sets the mechanical work of each joint
as low as possible, allowing joint actuators to be designed in parallel and still yield
a power-optimal manipulator. Power is balanced between all actuators such that any
differences in speed or torque for a single actuator can be adjusted from the motor’s
perspective by an appropriate gear reduction [122, 127, 143]. This removes the curse of
dimensionality from large power-optimization problems, decomposing it into kinematic
and actuator subproblems. This is in contrast to minimizing over-all power requirements
 5.1 efficiency metrics 47

Total Power Normed Net Power Antagonistic Power

( (
3

1 1 2 1
1 2
1
0 0
1
1
2 2
1
0
1
2
3 3 2
3

Σ|Pi | |ΣPi | 0

Sum of Squares Squared Net Power Power Quality

3
1 1 1 2

3 4 1
1 2
0 0
0 -1
1
2 -2
1
3 -3
4 4 3 2

Σ(Pi )2 (ΣPi )2 0

Figure 17: Antagonistic work compared to power quality in 2d. Each is a scalar field with level sets
drawn as contour lines. In 2d, power quality is a saddle centered at zero Q = 2P1 P2 .
Maximizing power quality brings the operational point closer to the 1-vector (primary
diagonal), reducing the chance that task distributions will require antagonistic work.
Note the mathematical and visual similarities between antagonistic work (the measure
of simultaneous opposite power) and power quality (the measure of power variance).

for the combination of possible kinematics with possible motor+gearhead pairs at all the
robot’s joints.
Knowing that contact forces for locomotion are predominantly along the leg-length, and
end-effector velocities are predominantly in the leg-swing direction, we can determine
the best kinematic designs without needing to touch an optimizer (Figure 19). The trivial
solution for minimizing antagonistic work is to make sure there is at least one motor per
limb that acts in the leg-swing direction only. Doing so guarantees that nominal contact
forces have no component on that actuator, making its power requirement zero. This is a
marginal power quality, however. So, all serially-connected limbs have at least marginal
power quality, and no antagonistic work in the nominal case. Simply considering the signs
of torques and speeds for different parts of the gait can show which designs have negative,
marginal, or positive power quality.
The idea of power quality has been extended into a power-optimal pseudoinverse which
is applied at run-time to minimize the amount of antagonistic work that a redundant
manipulator performs [25].
 5.1 efficiency metrics 48

Quantity Value Units

Resulting gait
Mass 75 kg
Stiffness 1.1 × 104 N/m Peak Force 1560 N
Apex height 0.93 m Stride Period 0.3578 s
Apex velocity 3 m/s Stride Length 1.036 m
Neutral length 1.0 m Average Speed 2.9 m/s
Touchdown angle 0.4 rad CoM Peak-to-Peak 0.0717 m
Gravitational constant 9.81 m/s2

Quantity Parallel Serial Spider Units

Nominal peak power 2.3 1.3 1.0 kW

Nominal MCoT 0.43 0.14 0.14 ·

Average MCoT 0.45 0.18 0.16 ·
StdDev MCoT ±0.08 ±0.04 ±0.03 ·

Mean power quality -2,600 0.00 150 ·

Task Parallel Serial Spider

3 m/s 0.4 m A
B A
75 kg 0.8 m

(a) Mechanism
0.6 m A 0.6 m

0.8 m
0.6 m B
0.6 m
B

(b) Nominal Joint Powers (c) Power Space, Quality

1.0 m Stride Length

Pnet 1.3 kW PA 2.3 kW PA 1.3 kW PA 1.0 kW

PB PB PB
110 J 330 J 110 J 110 J
0 0 0 0
-110 J -330 J -110 J -110 J

0s 0.36 s 0s 0.36 s 0s 0.36 s

0s 0.36 s
PB PB PB

Nominal
PA PA PA
Perturbed
Power Quality > 0
Power Quality < 0
3.5 kW

3 kW

3 kW

Average Cost: Average Cost: Average Cost:

340 ± 60 J 140 ± 30 J 120 ± 25 J
3.5 kW 3 kW 3 kW

Figure 18: Comparing three leg designs performing the same spring-mass running task. In order
to evaluate just the kinematics, none of these designs have components which store
energy (springs, flywheels, etc). The parallel mechanism has the highest peak power
requirements, does more work than the task requires, and therefore has the worst power
quality. The serial and spider designs do no antagonistic work nominally, but the spider
design requires less work for perturbed trajectories (a fact which is reflected in its higher
power quality).
 5.1 efficiency metrics 49

v v v
P2 = 0
f f f
P2 < 0 P1 > 0 P2 < 0 P1 > 0 P1 = 0

Figure 19: Decomposing force and velocity onto motor bases. Knowing that locomotion forces are
predominantly directed along the leg-length (towards the CoM) and that end-effector
velocities (for the actuators only) are predominantly perpendicular to the limb, we can
differentiate good kinematics from bad by projecting the task force and velocity onto
the actuators’ basis vectors. Task force is projected onto joint torques by drawing a
perpendicular line to each joint’s basis. Task velocity must be a linear combination of
joint bases, so the task velocity projects onto each basis using lines parallel to the other
joint’s basis (which is only true for invertible Jacobians). The resulting pair of joint
torque/speed shows whether the power requirement is positive, negative, or zero. Since
this task requires nominally zero power (task force and velocity are orthogonal), joint
powers must cancel.

5.1.2 Matching SLIP

In addition to its other benefits, the SLIP model gives us good heuristics for maximizing
efficiency. Bouncy, spring-mass trajectories solve the problem of antagonistic work during
level walking [8, 81], and the inclusion of passive compliance alleviates the problem of
cyclic work. Just like antagonistic work measures the amount of simultaneous, opposing
powers, cyclic work measures the amount of energy that is injected into a task just to
be removed at some point in the future. Cyclic exchange between gravitational potential
energy and spring potential energy typifies spring-mass locomotion. The efficiency of the
passive springs allows that gait energy to remain in mechanical form, not dissipated to
heat by motor braking.
Practical engineering intuition, research in dynamic locomotion, and animal physiology
all agree that series compliance is an important part of any limb design intended for
locomotion [64], but where and how is that compliance supposed to be added? The
standard approach is to use series-elastic actuators [114], which have compliant elements
directly in series with actuators. However, this may lead to an unnecessary number of
 5.1 efficiency metrics 50

springs in the system. In practice, springs can be placed anywhere in an articulated

mechanism and lead to a broad array of nonlinear functions at the end effector.
We can match the compliance of the SLIP in robot designs by tuning stiffness to match
that of the reduced-order model (thereby matching the gait’s resonant frequency) [83], and
making sure the directional behavior [85] of the task-space compliance is directed along
the leg length (just like the SLIP) [2].
Compliant force is modeled as the gradient of the compliant limb’s spring-potential well,
and the force generated by the springs is equal to the negative gradient of this potential
well at different deflected positions of the end-effector. Over large deflections, the well
can have various nonlinear shapes, but locally, this well is modeled as a simple parabolic
bowl and the force due to deflection can be modeled as a symmetric matrix in task space.
Visually, the force-deflection behavior is an ellipsoid whose axes are eigenvectors and
diameters are eigenvalues. So, matching the SLIP is a matter of tuning spring constants
until an axis of the compliance ellipsoid is directed along the long axis of the limb. Doing
so ensures that lengthwise forces result in lengthwise deflections, just as in the SLIP.
The potential well is a scalar field Vs (q) measuring the spring potential of a mechanism
deflected by ∆q from some set point q0 . Naturally, the set point q = q0 represents the
zero of the well, since this point corresponds to zero deflection in the springs. When
a force is applied, the equilibrium position for the mechanism corresponds to a point
q∗ = q0 + ∆q such that ∇Vs (q∗ ) = J⊤ f. This well-known relation comes from the Euler-
Lagrange equation using only a spring-potential term in the action L.
Equilibrium positions in arbitrary potential wells can be costly to compute, but quadratic
wells give good local approximations. As such, we can replace the spring potential with
the second-order Taylor series expansion around the set point,

∂Vs 1 ⊤ ∂2 Vs
Vs (q∗ ) = Vs (q0 + ∆q) ≈ Vs (q0 ) + ∆q + ∆q ∆q (51)
∂q q0 2 ∂q2 q0
1 ⊤ ∂2 Vs
≈ 0 + 0 + ∆q ∆q (52)
2 ∂q2 q0
 5.1 efficiency metrics 51

Because the set point is a minimum for the potential field, the zeroth- and first-order terms
vanish. With this local approximation, the equilibrium deflection of a mechanism becomes
the deflection ∆q that satisfies the equality

1 ⊤ ∂2 Vs ∂2 Vs
 
∇ ∆q ∆q = ∆q = J⊤ f, (53)
2 ∂q2 q0 ∂q2 q0

which is simply the equality of restitution and applied forces. When rearranged using
K−1 = C and J∆q ≈ ∆x (first-order approximation of f(q)), we find the task-space
compliance tensor, which relates applied loads to deflections,

K∆q = J⊤ f (54)

∆x = JCJ⊤ f. (55)

The tensor JCJ⊤ represents the approximate force-deflection relationship surrounding

a neutral position for the mechanism. Because it is a symmetric matrix, the task-space
compliance is orthonormal and can be visualized using ellipsoids whose major and minor
diameters are the eigenvalues of the matrix. Doing so allows for immediate feedback of the
compliant behavior of a design, as well as the ability to design for specific force-deflection
responses.
Task-space compliance can also be understood as the sum of symmetric dyads (a dyadic)
mapping force vectors to deflection vectors

[e1 , e2 , . . . ] = J columns of the Jacobian (56)

τi = e⊤
i f joint torque from task force (57)

∆θi = τi /ki joint deflection (58)

X
∆x = ei ∆θi first-order task deflection (59)
ei e⊤
X 
i
= f sum of dyads (60)
ki
||ei ||2
X 
⊤
= êi êi f normalizing bases (61)
ki
 5.2 manipulability metrics 52

e1,k1

Σ sca
le
= e2,k2

t
jec
pro
ei,ki Δxi
f Δx
f

Figure 20: Constructing compliant behavior from the sum of joint dyads. The task force is projected
onto each joint basis, then scaled by the squared length of the basis and the joint
||e ||2
stiffness, kii . Each component of the deflection ∆xi is added to the total deflection ∆x.
Projecting a unit circle of task forces produces the compliance ellipse.

There is one dyad per compliant degree of freedom. Geometrically, each dyad takes the
projection of the task force onto that joint’s basis vector in the task space, then scales the
projection by the length of the basis and the stiffness of the joint (Figure 20). The total
task deflection is the sum of these scaled projections of the task force, and the sum of
dyads itself is the task-space compliance JCJ⊤ . Understanding the dyad representation
gives much better intuition for the behavior of the mechanism by virtue of its geometric
properties and the ability to "visually compute" force-deflection mappings. This also
proves that serial mechanisms with hip and knee elasticity can never be tuned to give axial
compliance, since the leg-length force does not project onto the hip actuator, meaning only
the knee dyad participates, always deflecting partially in the leg-angle direction.
Physical compliance can be incorporated into Jacobian-transpose control [28], where the
task-space proportional gain tensor is the sum of physical stiffness and controller stiffness,
thus allowing for software-controlled impedance augmented with physical springs. Such a
method combines impedance control [132] with inverse kinematics [24, 74] in a way that is
appropriate for dynamic locomotion.

5.2 manipulability metrics

What "manipulability" means depends on how the controller is going to be interacting

with the hardware design. It is a measure of how well a machine can translate control
signals (in this case, electric motors as torque sources) into changes in a control target.
 5.2 manipulability metrics 53

Different control targets necessarily lead to different mechanical designs, so we need to

specify a reasonable control target for locomotion in general.
As discussed in Section 4.1, locomotion in general reduces to the control of contact force
(equivalently the acceleration of the robot’s mass-center). So, measuring the manipulability
of a machine reduces to measuring the limits on contact force and CoM acceleration.
Since locomotion is fundamentally mass transport, and eliciting changes in a mechanical
system’s motion necessarily requires center of mass velocity changes, we can say that
the all abstract locomotion controllers act on centroidal impulse and acceleration† (or
equivalently, the net contact force with the world). Control authority for a system with
such a controller can be measured as the maximum acceleration it can achieve for its center
of mass.

5.2.1 "Manipulability" for locomotion

Kinematic and Dynamic Manipulability are classic metrics for fixed-base manipulators, de-
scribing the robot’s kinematic and dynamic ability within the task space [86, 97, 159]. Peak
force [101, 106], velocity, and acceleration [29] in the task space are directional quantities; a
robot can be stronger or quicker in one direction than another. These metrics can be used
to measure the robot’s directional limit at the end effector [30], or they can measure how
"isotropic" [90] or uniform is the robot’s directional behavior‡ . Additionally, manipulability
can be used as a global performance index to grade a manipulator independently of its
pose [52, 103]. Locomotors are different in that they have multiple end effectors, and it is
actually the movement of the floating base that we care about§ . Additionally, we care about
how effectively the system can be reconfigured in anticipation of future locomotion needs.

† That is to say, all locomotion controllers reduce to eliciting CoM acceleration in one way or another. If they
did not, the robot would have no control over its own motion through the world. Because of this, we are
concerned with the controller’s ability to command CoM accelerations, whether or not that is a direct control
target of a particular algorithm. More CoM control authority necessarily leads to better manipulability of the
hardware as it pertains to system transport.
‡ Manipulability is sometimes regarded as a distance measure from singularities, since the isotropy of a
manipulator will collapse to zero at degenerate configurations.
§ Since the robot cannot modulate its momentum without contact, no contact necessarily implies no manipula-
bility.
 5.2 manipulability metrics 54

We can define the "manipulability" of a locomotor as its CoM acceleration limit. Accel-
eration is limited because electric motors are limited in how much torque they exert, and
it is directional due to mechanism kinematics (the robot could be stronger in leg extension
than in leg swing). This measure is normalized relative to the acceleration of gravity (i.e.,
measured in "g’s"). Since CoM acceleration depends on contact, and limbed locomotion
requires contact configuration to change frequently and in response to pushes or slips, we
must also consider the ability for the robot to control its end-effectors in space. We can use
the idea of Dynamic Manipulability to measure the maximum end-effector acceleration in
different configurations. End-effector peak acceleration is normalized by step length times
squared cycle frequency.

5.2.2 CoM acceleration limit

Calculating the acceleration limit for a floating-base robot in contact with the world is not
trivial, since contact models are ill-defined. Given the best case of rigid contact and zero
velocity, we can evaluate the acceleration limit of the system’s CoM by first evaluating the
maximum contact force

 −1
f = − JM−1 J⊤ JM−1 Bu. (62)

Using this equation, we project the joint-space hypercube of torque limits into the contact
force polytope (Figure 21). We can apply friction cone constraints by "trimming" the
contact force polytope by the friction cone geometry |fT | 6 µ|fN | (matlab has the function
polybool for this). After scaling the force polytope by system mass to leave acceleration,

we can shift the new acceleration polytope by the acceleration of gravity. We are left with
the maximum acceleration the system can achieve for its mass-center in all directions. Non-
zero joint velocities translate this polytope by Coriolis and velocity-product accelerations,
but its shape remains unchanged.
Considering the acceleration limit automatically balances reflected inertia vs total mass
of the robot. Too low of a gear ratio means the acceleration is sub-optimal because the
 5.2 manipulability metrics 55

u2

u1

u3

Figure 21: Computation of acceleration limits from actuator torque limits. The end-effector acceler-
ation limit transforms as in ẍ = JM−1 Bu and the the CoM acceleration limit transforms

−1
as in − JM−1 J⊤ JM−1 Bu/m, where m is the total system mass. The CoM limit
additionally must be trimmed by the contact friction cone and shifted by 1 g.

output torque of the actuators is not high enough to accelerate the robot’s mass. Too high
of a gear ratio means that spinning up the motor’s reflected inertia requires too much
torque, thus decreasing peak robot acceleration. This idea is analogous to the inertia ratio†
for single-DoF joints.
We do not necessarily want to maximize the acceleration potential of a locomotor. When
the system includes compliance (which all real systems do), having a 1:1 inertia ratio can
lead to larger tracking errors and more overshoot/ringing. There are good reasons for
intentionally including substantial compliance in a locomotor, so we want just enough
acceleration for walking, running, jumping, and whatever else the robot is designed to do,
but we still want to have as little reflected inertia as possible.

5.2.3 End-effector acceleration limit

Locomotion requires more than instantaneous centroidal acceleration; the system must
be able to reconfigure itself in preparation for future contact. This is also important to
position sensors, grasp an object, or prepare for future contact during a fall.

† Inertia ratio is a standard "rule of thumb" for specifying the ideal gear reduction for a motor driving an inertial
load. This measure applies to single degrees of freedom and, as the name suggests, is the ratio between the
motor’s reflected inertia and the load inertia. For an ideal, rigid system, the inertia ratio should be identity,
but flexible systems require a lighter reflected inertia than the load. The squared ideal gear reduction is equal
to the ratio of motor to load inertia, producing a maximal peak acceleration for the motor/load pair. A lower
gear ratio causes the motor torque to saturate due to the load inertia, while a higher gear ratio causes it to
saturate due to reflected motor inertia.
 5.2 manipulability metrics 56

We can measure the reconfigurability of a robot by computing the dynamic manipulabil-

ity of its end effectors relative to the robot’s root body

ẍ = Jq̈ + J̇q̇ task acceleration (63)

ẍ = JM−1 (Bu − h) + J̇q̇ joint-space forcing (64)

ẍ = JM−1 Bu controllable term (65)

which can be used to generate a task-space polytope of the acceleration limit given the joint-
space torque limits on u. Computing this polytope is as simple as evaluating Equation 65
for a all the positive and negative torque limits in joint space (can be generated as a Gray
Code in n bits, where n is the number of actuators). The linear projection of a hypercube
is guaranteed to be convex, so the acceleration limit polytope is the convex hull of the set
of values of ẍ. Since the uncontrollable terms J̇q̇ and JM−1 h are independent of u, they
will only translate the acceleration limit polytope, not affect its shape. In some cases, we
care about the directional dynamic manipulability, but we can reduce the polytope into a
simple scalar by finding the least-maximum acceleration. This scalar represents the radius
of the largest sphere we can inscribe within the polytope.
We can normalize the end-effector acceleration in the simple case of sinusoid trajectory
tracking, which is applicable to walking and running gaits. For the end-effector to travel a
distance L along a line in task space at frequency F, it requires a peak task acceleration
π2 2
of 2 LF . So, dividing the peak task acceleration of the manipulator by that factor gives
multiple of length-root-frequencies the manipulator can accommodate. A reasonable
π2 2
normalization for jogging at 2 m/s would be 2 (2 m)(1 Hz) = 9.87 m/ s2 , which is
interestingly very close to 1 g. For a normalized peak limb acceleration of 4, the limb can
approximately handle a stride length 4 times longer than given or 2 times as frequent (not
considering the workspace or motor speed limits of the machine).
 5.3 robustness metrics 57

5.3 robustness metrics

We need to try and create designs which do not fail when the world is not exactly the way
the robot thinks it is. This is not the same as component fatigue and mechanical failure,
but a failure to maintain stability of a pose or gait when disturbed. A robot’s hardware/-
software combination must be robust to unexpected changes in ground composition and
contact state. If the robot expects there to be a foothold where there isn’t, the controller
must "fail gracefully", performing autonomous IK to push the limb in the direction of
contact. The robot cannot rely on contact sensing in the timeframe where the control
system has little or no control over the robot’s dynamics (passive dynamics dominate).
Even more, there is no meaningful way to define "contact" for granular media, loose rubble,
and other non-rigid surfaces. As such, passive dynamics play a large role in the control
structure, since they define how the robot responds when disturbed or when the real
state of the robot is different from the measured state. Quite a lot of the robustness of a
system is determined by the controller, but there are situations which only the hardware
can manage, namely impacts and uncertain contact.

5.3.1 Impact inertia

The inherent, repeated inelastic collisions during locomotion can be a problem for robots
not designed to minimize impact mass. The intuition is clear: hammers have a substantial
mass in order to generate large, impulsive forces, while cheetahs, cursorial birds, and
many other animals have light, forward-facing tarsi and foot pads to minimize impacts [9].
If a leg is heavy or has a large unsprung mass, these collisions amount to a substantial loss
of kinetic energy [64]. If gearheads are subjected to these collisions, the impact forces can
be damaging. As such, it is important to shield inertias from the touchdown impact. We
analyze the impact inertia as the directional, projected mass of the manipulator into task
space (and simply normalized by the total system mass, leaving a percentage of unsprung
mass).
 5.3 robustness metrics 58

The real world is a complex place, and locomoting through it means substantial impacts
are going to occur. Since the robot will not always know about impending impacts, we
must design the hardware system to minimize the force and impulse due to inevitable
collisions between its end effectors and the world. Reducing impact forces reduces the
chance of damaging components, and reducing impulses limits the discontinuous effect of
impacts on the state of the robot, thereby reducing the control burden.
We may not be able to control the mass of objects in the world, but we can control how
heavy limbs are. This is not just reducing over-all weight, but reducing the operational
inertia of the end-effectors. Operational inertia is a function of the spatial inertia of the
robot’s links as well as reflected inertia at its joints. As is the case with large gear reductions,
the joint reflected inertia can easily dominate the operational inertia of a limb.
Passive compliance can help reduce the effects of impact by limiting the operational
inertia to distal masses only and buffering the transmission of force to the upstream
components. Compliant distal joints work well for buffering impacts as well as recycling
gait energy (an efficiency bonus). Additionally, compliant pads on the end-effectors are
effective at further reducing impact loading, although they typically are viscoelastic and
do not recycle energy.
We can derive a machine’s impact inertia by assuming that acceleration forces will
dominate during an impact. The resulting impact equation for a manipulator is

M∆q̇ = J⊤ Λ (66)

J∆q̇ = JM−1 J⊤ Λ (67)

 −1
JM−1 J⊤ ∆ẋ = Λ, (68)

where ∆q̇ is the change in joint velocity for an impulse H, and J is the Jacobian matrix
at the point of impact (see [2]). Equation 68 shows the projected impact mass in task
space, relating instantaneous velocity changes to impulses. The matrix term is symmetric,
representing an ellipsoid in task space which shows the directionality of impact inertia.
Projecting the joint-space inertia onto different points on the limb easily visualizes
the directionally-dependent impact mass at those points [2, 14]. This leads to the simple
 5.3 robustness metrics 59

Δv
Λ
Figure 22: Relationship between uniform ∆v and impulse Λ for a homogeneous bar. The impact
inertia is defined at every point on any rigid body, whether free-floating or attached
to an articulated body. Larger principle axes of the inertia ellipsoids represent heavier
directions and shorter principle axes represent lighter directions.

identification of "light" and "heavy" directions for any point on a limb (Figure 22). One
simple result that agrees with literature and biology is the forward-facing tarsus, which
receives impacts more perpendicular to its length [2, 69].
This method respects the presence of elastic elements and transparent actuators, and
therefore gives accurate visual feedback of the true impact response of a machine. As such,
it is easy to see which actuators or degrees of freedom need compliance to buffer impacts
and which ones don’t. For example, high-transparency actuators (those with very low
reflected inertia) do not contribute significantly to the impact ellipse and therefore do not
need series compliance to reduce force spikes.
One necessary assumption to ensure the validity of this method is that we must make
sure that the time period of impacts is far shorter than the natural period of the suspension
system. In reality, stiff springs will transfer the ground reaction force to massive compo-
nents almost as quickly as a "rigid" element would (there is no such thing as rigid in reality,
just high stiffness). Impact inertia considers the very instant of impact, so any modeled
springs, no matter how stiff, will buffer the upstream mass, so very stiff springs will give
inaccurate results. This is because springs of any stiffness in the model are still much softer
than the infinite stiffness of rigid links.
Minimizing the magnitude of the operational space inertia is effective at reducing impact
effects, but it is not a well-normalized metric. The Impact Mitigation Factor is a recent
unitless measure of backdrivability during impact, allowing us to compare robots of
different sizes, weights, and joint types [154].
 5.4 feasibility metrics 60

5.4 feasibility metrics

Feasibility metrics are perhaps the most important, because they determine if the design can
even meet task requirements. While constraints are common in numerical optimizations,
it generally does not make sense to implement feasibility checks as hard constraints.
Rather, we can take another common technique from computer optimization and add
penalty functions as objectives [156]. Penalty functions tell us how far a design is from being
feasible, giving us greater feedback on different designs. Hard constraints can apply when
generating designs, where it is easy to say that links can be no shorter than x, link density
must be at least ρ, the legs must have a total length ℓ, and so on.
Feasibility is about quantifying the degree to which robots have to violate physical
constraints in order to follow a task. It may sound pedantic, but chairs are perfectly
efficient and fairly robust to different methods of being sat on, but they are terrible
locomotors and cannot carry anyone around (yet). Specifically to locomotion, a locomotor
must be able to transport itself through the world through unilateral, friction-constrained
contact while obeying motor torque/speed limits. If it cannot, then it is infeasible.
When designing a machine, we typically do not have hard constraints unless they are
trivial to enforce (such as link length). It is better to know how much a constraint is violated
in a design and use that information to migrate towards a more feasible design.
Measuring "how much" a constraint has been violated is exactly the purpose of penalty
functions (which are not to be confused with barrier functions). The most basic penalty
function is the quadratic loss, where the penalty for violating an equality constraint is
simply the square of the error (Figure 23). Violating an inequality constraint results in the
same quadratic loss, but unviolated inequality constraints carry no cost

φeq (e) = e2 (69)




0, e 6 0

φineq (e) = (70)

e2 , e > 0


 5.4 feasibility metrics 61

invalid

invalid
valid valid

e e
Equality penalty Inequality penalty

Figure 23: Rather than enforcing hard constraints, the quadratic loss penalty function encodes "how
much" an equality or inequality constraint is being violated.

When trying to optimize or follow a trajectory, we can include a physics violation cost

Z 2
Lphysics = J⊤ f + Bu − Mq̈ − h dt (71)

which accumulates the error between system dynamics and those required by the trajectory
[99].
Friction cone and motor speed/torque limits are unilateral constraints, but they can be
handled in the same way. The constraints

|fT | 6 µfN tangential force limit (72)

R
kb |θ̇| + |τ| 6 Vmax winding voltage limit (73)
kb

become the penalty functions

Z
Lcone = max(0, fT 2 − µ2 fN 2 ) dt (74)
R
Z
Lvoltage = max(0, kb |θ̇| + |τ| − Vmax )2 dt (75)
kb

such that Lcone and Lvoltage are the feasibility metrics for a particular gait according to
friction cone and torque-speed limitations.
 6 METRIC-DRIVEN DESIGN PROCESS

Using several unrelated metrics prevents over-specialization of designs, but they need
to be combined using some exterior structure. For this, we can use the notion of Pareto
optimality [27], where a design "dominates" another only when it is better according to
every metric defining the objective space. If designs do not dominate each other, then they
are considered equal and merely make different trade-offs between their metrics. This
method effectively creates a singular, weakly-ordered metric by combining two or more
existing metrics.

6.0.1 Corner sort

The multi-objective space has a necessary weak ordering, so members cannot be trivially
compared. Pareto-optimality tells us that there exists a set of non-dominated solutions
which are better in at least one respect than every other member. In order (no pun
intended) to rank designs in a multi-objective space, we can use the Corner Sort algorithm
to iteratively find the non-dominated set, assign a rank, then find the non-dominated
set of the remaining dominated solutions [152]. This partitions the designs into "shells"
of increasing rank, allowing any pair to be meaningfully ordered even if one does not
completely dominate the other (Figure 24).

6.1 automated design generation

We can partially automate candidate design generation by stochastically searching the

design space [140]. This is no Grand, Unified Optimizer‡ , but it works sufficiently well for

‡ It is a very exciting idea to have one single optimization program which translates a task specification into a
hardware design and a corresponding optimal controller, but that goal is well out of the scope of this work.

62
 6.1 automated design generation 63

Figure 24: Corner sort algorithm (graphic adapted from [152]). All solutions begin unranked
(rank = 0). Keeping track of iterations i, do the following: 1) unmark all unranked
solutions, 2) find the unmarked corner solutions, set their rank to be i, mark them, and
mark all solutions they dominate, 3) repeat at 2 until all solutions are marked, 4) repeat
at 1 until all solutions are ranked.

helping direct the design process. Stochastic search is a simple pattern that uses modular
components (Figure 25):

1. Factories for randomizing continuous parameters of given template designs.

2. A population of instantiated designs.

3. A set of metrics which are evaluated for each member of the population.

4. Corner sorting algorithm for providing the Pareto rank of the members.

5. A selection scheme for reducing the population (we cannot exhaustively search the
design space).

6. A "temperature" schedule for determining how much to randomize each member

of the population when replacing culled members. High initial temperatures allow
the population to cover the design space, and slowly-cooling temperatures allow the
population to refine towards the true Pareto frontier.

We are dealing with trade-offs, so we need a sufficiently large population which can cover
the Pareto frontier in the design space (which is projected into the metric space). The
design space is sufficiently complicated that we do not want to risk falling into local
optima by using hill climbing. Using just a single candidate with random hill climbing is
practically guaranteed to be suboptimal, and it gives us no information about the trade-offs
that can be made in the metric space.
 6.1 automated design generation 64

Layer 0

Pareto-Optimal
SPECS OUT Design Set
ction
Sele
Corner
Sort

Templates Factory
Randomized M1
Candidates
Parameters:
{a1,a2,a3,...}

Design
IDEAS IN Factory
Evaluation
Parameters:
{b1,b2,b3,...}
M[0] = costOfTransport(trial);
M[1] = impactInertia(trial); M2
M[2] = compliance(trial);
Factory .
.
Parameters:
.
{c1,c2,c3,...}

Figure 25: Diagram of the iterative design process. Topologies are passed in as templates with
simple parameters such as link lengths and gear ratios. Each template has a corre-
sponding factory which knows how to stochastically vary that template’s parameters.
A population of candidate designs is evaluated according to a set of metrics, mapping
the design space into the objective space. The corner sort algorithm finds the "layers" of
non-dominated candidates in the population, and binary tournament selection removes
dominated designs while maintaining diversity. This process is both a formal multiobjec-
tive optimization algorithm and an analogy for practical iterative design processes used
by engineers.

6.1.1 Template designs, factories

We can use human creativity to generate a handful of different design templates, each
with a different link topology, motor locations, and so on. These templates have numerical
parameters for link-lengths, gear-reductions, motor modules, and so on that are trivial to
vary automatically within a set of allowed values.
Factories randomly generate candidate designs that meet design plausibility constraints
such as nonzero link length, nonzero mass, etc. For the stochastic search algorithm, the
factories produce randomized designs from existing designs until the population has
grown by some factor (usually double).

6.1.2 Binary tournament selection with redraw

We want to limit the population of designs while also maintaining diversity, so we need
an appropriate rule for removing members. There are many heuristics for this [51], but
 6.1 automated design generation 65

binary tournament selection is a simple rule which does not limit diversity† . BTS is as simple
as drawing two designs randomly from the sorted population and removing the one with
lower rank (redraw is allowed, for simplicity). Do this until the population size has been
reduced by some factor (usually half).

6.1.3 Simulated annealing

We can use a temperature schedule to gradually decrease the amount of randomness in

the design factories [84]. This amounts to large steps through the design space during the
first iterations, attempting to rapidly approach the Pareto frontier, and small steps during
the later iterations (gradually refining the approximate frontier).
We can also incorporate the annealing temperature into the binary tournament selection
algorithm, keeping the worse design with some probability. This helps maintain a diverse
population at high temperatures.

† We could just remove all the dominated solutions from a population, but that decreases coverage of the design
space. We want to leave as many varied designs on the table as we can while still hunting for the Pareto
frontier.
7 EXAMPLE DESIGNS

To illustrate the use of locomotion metrics, we evaluate select metrics for three designs: a
simple 2-link serial mechanism, an ATRIAS-like parallel mechanism, and a polar mecha-
nism. These designs all have invertible kinematics (neglecting spring deflection), but all
the metrics presented operate on any general combination of redundant serial and parallel
kinematics (including bi-articulation). See Section 3 for details on modeling.
In a complete design process, the results of these evaluations would "feed back" into the
next iteration of designs. Designers would adjust link lengths, shift mass, choose different
motors, and so on to try to improve the next round of metrics. This iterative design process
teases us with the potential for automation via stochastic search. Section 6 discusses such
a method, but that method will not be followed here, since we are focusing on the metrics
themselves.
Each design can be evaluated using the visual form of numeric metrics to get a feel
for how well a design will behave to the point of sanity-checking numerical values or
optimization results (Section 5). This is a transferable technique for guiding the design
process when creatively generating new designs. Visual representations of different limb
properties can be quickly and easily interpreted even with little or no understanding of
the underlying mathematics.
The example designs we evaluate here are constructed using uniform-density bars for
links, and all designs use the same density material in their construction. As such, the
inertial properties of the links are derived from the area of "ink" used to draw them.
Actuators are represented by additional masses at the actuated joint and the associated
reflected inertia in the system’s mass matrix diagonals. All designs have a leg length of 1
meter, use the same motors (with possibly different gearhead reductions), and have series
compliance tuned for the same nominal toe stiffness. Each leg design is attached to an
identical floating base. Even though these designs are made out of the same "stuff", they

66
 example designs 67

are used as an exhibition of metrics only and not compared in detail, since real designs will
be significantly different in their weight distribution. Table 1 shows the collected numerical
evaluation of all metrics and Figures 26, 27, and 28 show the visual representations.
We evaluate each designs’ CoM acceleration limit to track their ability to generate motion,
their end-effector acceleration limit to track their ability to reconfigure, their compliant
behavior as appropriate for spring-mass locomotion, their impact mass (smaller is better),
and their power quality (positive is better).

com acceleration limit This polytope represents the robot’s ability to change
its motion when in contact with rigid ground (making it the "manipulability" of floating-
base manipulators). We draw a convex envelope around the robot’s mass-center to show
the boundary of possible accelerations the robot can achieve given motor torque and
friction-cone constraints. CoM acceleration limits are directional, so a robot may be able
to accelerate straight up more easily than it can accelerate laterally (represented by a
thin and tall polytope). All floating-base robots on Earth will have the same downward
acceleration limit: 1 g (unless it can pull itself toward the ground using a bilateral contact
constraint). We are not always interested in the isotropy of manipulability; many more
fitting measures could be drawn from the acceleration envelope such as vertical limit while
standing, forward limit in a crouched position, lateral limit in a split-leg pose, and so
on. We primarily affect the CoM acceleration limit by changing gear ratio, since this will
balance reflected inertia of the actuators against the total mass of the robot.

end-effector acceleration limit Just as the CoM has a directional acceleration

limit, the robot’s end-effectors also have such a limit. Here, we consider this acceleration
relative to the robot’s root frame (making this metric very much like the Dynamic Ma-
nipulability of a fixed-base manipulator). The end-effector acceleration is also directional,
where the envelope’s distance from the end-effector represents a higher peak acceleration
in that direction. General, directionally-uniform behavior is not necessarily the goal, since
the acceleration requirement for locomotion is very much anisotropic and configuration-
dependent. While manipulability is independent of the task, the project/task requirements
 example designs 68

necessarily determine different "cross-sections" of manipulability which matter more than

others (the anisotropies).

end-effector compliant response Compliant behavior may be difficult to grasp

when thinking in terms of individual joint stiffnesses, but the whole-limb behavior is
succinctly represented as an ellipse centered at the end-effector. This ellipse is the result
of a circle of unit forces being scaled along the principle axes of the ellipse (major and
minor diameters). Each unit force on this circle maps to the scaled point on the ellipse
representing the deflection of the end-effector under that force. Since this is a scaling
operation, the directions of forces along the major and minor diameters remain unchanged.
So, a unit force along the long axis of the ellipse generates a large equilibrium deflection
in that direction, and that same force along the short axis results in less deflection. For
spring-mass locomotion, we desire leg-length forces to result in leg-length deflections, so
we want to tune the leg designs to have a principle compliant axis along the leg length at
the desired spring rate.

end-effector impact mass Impact with rigid ground is modeled as a linear re-
lationship between the end-effector’s velocity delta (before and after contact) and the
impulse required to generate that delta. Just like compliance, this relationship is directional
and can be visualized as an ellipse with major and minor axes. Long axes represent a
heavier impact mass (more impulse to elicit a unit velocity delta in that direction), and
short axes represent a lighter impact mass.

power quality We can use a simple task definition to evaluate how well different
designs decompose the net power required onto individual actuators. It is important
to limit actuators from contributing simultaneous positive and negative powers due to
the lack of regeneration and larger power requirements. Visually, task velocity and force
decompose onto the design’s actuator basis vectors (lines representing the direction/speed
of the end-effector given unit velocity each actuator, fixing all other actuators). Force
projects literally (as a scalar product with each basis), and velocity projects as the sum of
 example designs 69

scaled basis vectors. Power quality can be coarsely-graded as positive or negative, or it

can be evaluated using the products of speed and torque for each actuator (area of the
rectangle with side-lengths of the force and velocity projections).

7.0.1 Polar Design

The first design (Figure 26) is simply a prismatic leg-length actuator pivoted by a revolute
leg-angle actuator and connected to an unconstrained torso. Each actuator includes a
serially-actuated spring to add compliance. This design has the same toe kinematics (e1 and
e2 basis vectors in the figure) as various straight-line linkages, and thus can represent the
compliance, power quality, and manipulability of any other revolute-prismatic design (but
not impact inertia). Due to symmetry, the compliant and impact response are symmetrical
around the leg-length, and so is the manipulability. Such a design has marginal power
quality, due to the decomposition of walking/running forces onto a single actuator alone,
meaning only one motor does any work (although this neglects inertial forces, which
certainly have a leg-angle component). It is trivial to tune the compliant response of
this design to match the SLIP, and achieving the desired stiffness is as simple as setting
the prismatic spring to exactly that stiffness. Since leg-length forces decompose onto the
prismatic spring only, walking/running forces will only ever elicit a leg-length response.

7.0.2 2-link Serial Design

The second example design has two links connected rotationally at the knee and hip.
Like the other leg designs, this one is connected to a floating base. Since the links are
connected serially, the hip acts to swing the leg as a whole, and the knee acts to swing just
the second link. Each actuator has a serially-connected spring at its output, so as to add
compliant behavior. The 2-link design is asymmetric, meaning it’s compliant, impact, and
manipulability behavior is asymmetric. There is more impact mass along the length of the
shin, due to the actuator located at the knee. Actuator 1 (which has e1 as its basis vector)
 example designs 70

G1 = 30:1
k1 = 1000 N.m/rad m = 33 kg
Tmax = 12.2 N.m

G2 = 85 rad/m
k2 = 11 kN/m
Tmax = 12.2 N.m
m
5
0.

g = 9.81 m/s2

(min, max)
(0.70, 2.12) g CoM acl
(11.5, 21.1) jog toe acl
0.5 m

(2, 11) kN/m compliance

(6.45, 9.93) % impact

e1

e2

Figure 26: The first leg design is a revolute-prismatic "pogo stick" design. The hip includes a gear-
motor and a series spring, and the leg-length is a sprung slide actuated by linear actuator
(let’s say it is a ballscrew driven by a rotary motor). Locomotion forces decompose onto
the leg-length actuator only, and swing velocity onto the leg-angle only, giving this
design a marginal (zero) power quality.
 example designs 71

G1 = 30:1
k1 = 1100 N.m/rad
Tmax = 12.2 N.m

G2 = 25:1
k2 = 1100 N.m/rad
Tmax = 12.2 N.m m = 33 kg

(min, max)

m
(0.69, 2.27) g CoM acl

5
0. g = 9.81 m/s 2 (8.11, 29.1) jog toe acl
(1.68, 11.5) kN/m compliance
(3.87, 23) % impact
0.
5
m

e2

e1

Figure 27: The second leg design is a 2-joint serial mechanism. Each joint includes a gearmotor and
a series spring. Just like the pogo stick, locomotion forces act at the knee only and leg
swing is controlled by the hip, giving the serial design a marginal (zero) power quality.
Unfortunately, since there is no series compliance along the length of the shin, the knee
motor mass participates in impacts directed along the shin. As such, the max impact
percentage is higher for this design than the others. Also, since leg-length forces do
not decompose onto the hip actuator, only the knee spring will ever deflect under load,
meaning this design can never be tuned for a leg-length principal compliance.

does not act in the leg-length direction, so walking/running forces decompose only onto
the knee actuator (e2 ). As such, this design can never have a compliant SLIP response,
since leg-length forces never generate leg-length deflections, no matter the spring constants
used at the knee and hip (you can use the dyadic representation of compliance to prove
this graphically).
 7.1 comparison 72

7.0.3 4-bar Parallel Design

Lastly, we have a pantograph leg with four equal-length links. Because this design has
a loop-closure, we project the open-chain dynamics of two, two-link chains using the
constraint that the ends of these chains share position, velocity, and acceleration (see
Section 3). The two series-compliant actuators are located at the hip and drive each of
the proximal links. As shown in Figure 28, the parallel design is highly symmetric in its
dynamic behavior. It is very good at accelerating its mass-center vertically (which was
one of the intuitive reasons for using a parallel mechanism in ATRIAS, although it was
just gut instinct at the time). It has isotropic impact and compliant responses in the pose
shown, but these responses scale along the length of the leg as the leg extends and retracts
(and retain a mirror symmetry about the length of the leg). The critical flaw with this
design, as described earlier, is the antagonistic work that must be done during walking and
running. Because leg-length forces decompose into equal and opposite actuator torques,
and leg-swing velocity decomposes into equal actuator speeds, there is nominal equal and
opposite work being done by the actuators. As such, this design has very poor power
quality.

7.1 comparison

Table 1 shows the collected evaluation of the metrics shown in the leg design figures. Since
most of these metrics are directional, the numerical quantities are taken as the maximum
value in any direction, along with an isotropy value (which is simply the least-maximum
value divided by the limiting value). Isotropy isn’t important if a particular project is not
concerned with the uniformity of a design’s dynamic behavior. Similarly, only some of
these metrics may be important, depending on the goals of different projects. For example,
fixed-base manipulators aren’t concerned with power quality, because they are attached to
wall power. If a design is focused on actuator transparency, then compliance becomes less
important.
 7.1 comparison 73

G1 = 30:1
k1 = 2750 N.m/rad
Tmax = 12.2 N.m

G2 = 30:1

m
k2 = 2750 N.m/rad

=
33
Tmax = 12.2 N.m

kg
m
5

g = 9.81 m/s2
0.
0.
5

(min, max)
m

(0.56, 2.07) g CoM acl

(17.2, 24.4) jog toe acl
(11, 11) kN/m compliance
(8.63, 8.63) % impact

e1 e2

Figure 28: The third leg design is a parallel mechanism with two series-elastic motors mounted on
the body. Two branches of links require a loop-closure constraint at the toe to enforce
the four-bar linkage kinematics. Because leg-length forces decompose equally onto both
motors, which also both have to rotate to swing the leg, the parallel design has a negative
power quality, indicating antagonistic work requirements.
 7.1 comparison 74

Three designs are too small a set to draw any general conclusions, but we will compare
these designs nonetheless. In reality, the results shown here would feed back into the
design process, directing modifications to these designs, or necessitating the addition
of completely new designs into the mix. Further designs could include straight-line
mechanisms, articulated mechanisms with more links (the metrics have no issue with
redundancy), 3D designs (using ellipsoids and polytopes with volume instead of area),
and so on.
Tuning gear reductions has an effect on impact mass (unless there is a series spring below
the high-reduction actuator) and trades off CoM acceleration for toe acceleration (just like
inertia matching for a single DoF system). The serial design has the best combination of
CoM and toe acceleration limits, but the isotropy is low at the toe. The CoM acceleration
isotropy is fairly low across the board due to the friction cone constraint on contact forces
(all robots can only accelerate at 1 g downward). Leg stiffness is tuned to the same value
for all designs, but the asymmetry of the serial design means it will never have pure SLIP
compliant behavior. The serial design has the worst unsprung mass along the length of
its shin, due to the actuator mass at the knee, but it actually has the lowest unsprung
mas perpendicular to the shin. Using the predominantly leg-length forces and leg-angle
velocities of walking and running, we see that the prismatic and serial designs have
nominally zero antagonistic work (marginal power quality), and the parallel design has
significant antagonistic work (negative power quality).
 7.1 comparison 75

Quantity Pogo Serial Parallel Units Order

CoM acceleration limit 2.12 2.27 2.07 g >

0.33 0.30 0.27 "iso" >
Toe acceleration limit 21.1 29.1 24.4 "jogs" >
0.55 0.28 0.70 "iso" >
Leg stiffness (tuned) 11.0 11.5 11.0 kN/m =
Unsprung mass 9.93 23.0 8.63 % <
0.65 0.17 1.00 "iso" >
Power quality =0 =0 <0 W2 >

Table 1: Properties of three leg designs. Metrics are evaluated at a nominal leg length of 0.707 m.
CoM acceleration limit is normalized by the acceleration due to gravity, toe acceleration
is normalized by the peak acceleration required by a 2 m/s jogging gait (hence "jogs",
9.87 m/s2 ), and impact is normalized as a the percent of robot mass which is involved in the
impact (unsprung). "Iso" is short for "isotropy" (the ratio of least to greatest measurement
for directional quantities), where 1 is completely uniform. The least-maximum value of the
metric is the limit times the isotropy value. These values are represented visually in the
following figures.
8 CONCLUSION

This work has been collected for lack of a detailed roadmap in mechanical design for loco-
motion. It is an intricate problem, rooted in dynamics and with close connections to control
design and contact mechanics. Such interconnection makes locomotion a particularly inter-
esting problem to analyze, because it resists established norms of controller-dominated,
just-needs-to-exist hardware designs in favor of "mechanical intelligence" and the teaming
of hardware and software.
Clearly defined task parameters and design goals guide us in our search for better
hardware designs for locomotion systems. Competing design goals lead to trade-offs in
the design space, and the iterative nature of design means that the only "solution" to the
design problem is continual refinement of existing designs.
We have attempted to render the given limb-design metrics in geometric forms such
that the behavior of the limb design is visually apparent. Compliance and impact behavior
are represented by ellipsoids indicating the deflection and impulse response for applied
forces and velocity deltas, respectively. Floating-base acceleration limits are inherently
convex polytopes (intersection of half-spaces) showing the directional limit of controlled
acceleration for a locomotion system. Power quality can be evaluated by drawing actuator
basis vectors in the task space and projecting the task’s force and velocity onto those
bases, indicating the sign of actuator powers and showing any antagonistic work. Some of
these processes can be largely performed on paper, allowing designers to roughly evaluate
candidate designs without any modeling or simulation.
A structured design process encodes what is fundamentally an optimization program,
where our design process is a human-in-the-loop algorithm designed to solve that program.
While what we have is no Grand, Unified Optimizer, the metrics can be evaluated in
simulation, and the design process is written in a way that is amenable to computer

76
 8.1 disclosure 77

automation. The largest hurdles are through-contact trajectory optimization with realistic
physics and limb topology grammars to encode and randomize design candidates.

8.1 disclosure

Andrew Abate is an employee of Agility Robotics, a company that licenses technology

derived from Andrew’s research at Oregon State University and sells versions of the Cassie
bipedal robot for locomotion research.

8.2 acknowledgments

People are products of their environment, therefore academics are products of each other.
This work is naturally a shared work of myself and all the smart, driven co-workers I have
had the pleasure to know. Namely my labmates at Oregon State over the years: Christian
Hubicki, Hamid Vejdani, Siavash Rezazadeh, Daniel Renjewski, Jesse Grimes, Mikhail
Jones, Brent Piercy, Kevin Kemper, Devin Koepl, Andrew Peekema, and Ryan Van Why. I
also deeply appreciate the time of my formal (and ad-hoc) advisors: Jonathan Hurst, Ross
Hatton, Matt Campbell, Burak Sencer, and C. David Remy.
This material is based upon work supported by the National Science Foundation under
Graduate Research Fellowship Program Grant No. 1314109, and also under DARPA award
number W911NF-16-1-0002.
BIBLIOGRAPHY

[1] Andy Abate. “Preserving the Planar Dynamics of a Compliant Bipedal Robot with
a Yaw-Stabilizing Foot Design.” In: (2014).

[2] Andy Abate, Ross L Hatton, and Jonathan W Hurst. “Passive-dynamic leg design for
agile robots.” In: Robotics and Automation (ICRA), 2015 IEEE International Conference
on. IEEE. 2015, pp. 4519–4524.

[3] Andy Abate, Jonathan W Hurst, and Ross L Hatton. “Mechanical Antagonism in
Legged Robots.” In: Robotics: Science and Systems (RSS). 2016.

[4] Jeffrey Ackerman and Justin Seipel. “Energy efficiency of legged robot locomotion
with elastically suspended loads.” In: IEEE Transactions on Robotics 29.2 (2013),
pp. 321–330.

[5] Mojtaba Ahmadi and Martin Buehler. “Controlled passive dynamic running exper-
iments with the ARL-monopod II.” In: Robotics, IEEE Transactions on 22.5 (2006),
pp. 974–986.

[6] Antonio Alba, Francesco Bucchi, Francesco Frendo, and Marco Gabiccini. “Kine-
matic Optimization of the Arm of a Working Machine.” In: ASME 2014 International
Design Engineering Technical Conferences and Computers and Information in Engineer-
ing Conference. American Society of Mechanical Engineers. 2014, V006T10A071–
V006T10A071.

[7] R. McNeill Alexander. Elastic mechanisms in animal movement. Cambridge University

Press, 1988.

[8] R. McNeill Alexander. “Three uses for springs in legged locomotion.” In: Interna-
tional Journal of Robotics Research 9.2 (1990), pp. 53–61.

78
 bibliography 79

[9] R. McNeill Alexander, MB Bennett, and RF Ker. “Mechanical properties and func-
tion of the paw pads of some mammals.” In: Journal of Zoology 209.3 (1986), pp. 405–
419.

[10] R. McNeill Alexander and Alexandra Vernon. “The mechanics of hopping by

kangaroos (Macropodidae).” In: Journal of Zoology 177.2 (1975), pp. 265–303.

[11] Emanuel Andrada, Christian Rode, and Reinhard Blickhan. “Grounded running in
quails: simulations indicate benefits of observed fixed aperture angle between legs
before touch-down.” In: Journal of theoretical biology 335 (2013), pp. 97–107.

[12] Ben Andrews, Bruce Miller, John Schmitt, and Jonathan E Clark. “Running over un-
known rough terrain with a one-legged planar robot.” In: Bioinspiration & biomimetics
6.2 (2011), p. 026009.

[13] Ömür Arslan and Uluç Saranlı. “Reactive planning and control of planar spring–
mass running on rough terrain.” In: Robotics, IEEE Transactions on 28.3 (2012),
pp. 567–579.

[14] Haruhiko Asada. “Dynamic analysis and design of robot manipulators using
inertia ellipsoids.” In: Robotics and Automation. Proceedings. 1984 IEEE International
Conference on. Vol. 1. IEEE. 1984, pp. 94–102.

[15] David Baraff. “Linear-time dynamics using lagrange multipliers.” In: Proceedings of
the 23rd annual conference on Computer graphics and interactive techniques. ACM. 1996,
pp. 137–146.

[16] DN Beal, FS Hover, MS Triantafyllou, JC Liao, and GV Lauder. “Passive propulsion

in vortex wakes.” In: Journal of Fluid Mechanics 549 (2006), pp. 385–402.

[17] JE Bertram, Andy Ruina, CE Cannon, Y Hui Chang, and Michael J Coleman. “A
point-mass model of gibbon locomotion.” In: Journal of Experimental Biology 202.19
(1999), pp. 2609–2617.

[18] Pranav A Bhounsule, Jason Cortell, and Andy Ruina. “Design and control of Ranger:
an energy-efficient, dynamic walking robot.” In: Proc. CLAWAR. 2012, pp. 441–448.
 bibliography 80

[19] Aleksandra V Birn-Jeffery, Christian M Hubicki, Yvonne Blum, Daniel Renjewski,

Jonathan W Hurst, and Monica A Daley. “Don’t break a leg: running birds from
quail to ostrich prioritize leg safety and economy on uneven terrain.” In: Journal of
Experimental Biology 217.21 (2014), pp. 3786–3796.

[20] Reinhard Blickhan. “The spring-mass model for running and hopping.” In: Journal
of biomechanics 22.11-12 (1989), pp. 1217–1227.

[21] Reinhard Blickhan, Andre Seyfarth, Hartmut Geyer, Sten Grimmer, Heiko Wagner,
and Michael Günther. “Intelligence by mechanics.” In: Philosophical Transactions of
the Royal Society of London A: Mathematical, Physical and Engineering Sciences 365.1850
(2007), pp. 199–220.

[22] Edgar Bolívar, Siavash Rezazadeh, and Robert Gregg. “A General Framework for
Minimizing Energy Consumption of Series Elastic Actuators With Regeneration.” In:
ASME 2017 Dynamic Systems and Control Conference. American Society of Mechanical
Engineers. 2017, V001T36A005–V001T36A005.

[23] Ben Brown and Garth Zeglin. “The bow leg hopping robot.” In: Robotics and
Automation, 1998. Proceedings. 1998 IEEE International Conference on. Vol. 1. IEEE.
1998, pp. 781–786.

[24] Samuel R Buss. “Introduction to inverse kinematics with jacobian transpose, pseu-
doinverse and damped least squares methods.” In: IEEE Journal of Robotics and
Automation 17.1-19 (2004), p. 16.

[25] Nathan M Cahill, Thomas Sugar, Matthew Holgate, and Kyle Schroeder. “Under-
standing Power Loss due to Mechanical Antagonism and a New Power-Optimal
Pseudoinverse for Redundant Actuators.” In: ASME 2017 International Design Engi-
neering Technical Conferences and Computers and Information in Engineering Conference.
American Society of Mechanical Engineers. 2017, V05BT08A077–V05BT08A077.

[26] Matthew I Campbell, Jonathan Cagan, and Kenneth Kotovsky. “A-design: an agent-
based approach to conceptual design in a dynamic environment.” In: Research in
Engineering Design 11.3 (1999), pp. 172–192.
 bibliography 81

[27] Yair Censor. “Pareto optimality in multiobjective problems.” In: Applied Mathematics
and Optimization 4.1 (1977), pp. 41–59.

[28] Chien-Chern Cheah. “Task-space PD control of robot manipulators: unified analysis

and duality property.” In: The International Journal of Robotics Research 27.10 (2008),
pp. 1152–1170.

[29] D-Z Chen and L-W Tsai. “The generalized principle of inertia match for geared
robotic mechanisms.” In: Robotics and Automation, 1991. Proceedings., 1991 IEEE
International Conference on. IEEE. 1991, pp. 1282–1287.

[30] Hee-Byoung Choi and Jeha Ryu. “Convex hull-based power manipulability analysis
of robot manipulators.” In: Robotics and Automation (ICRA), 2012 IEEE International
Conference on. IEEE. 2012, pp. 2972–2977.

[31] Steven H Collins, Peter G Adamczyk, and Arthur D Kuo. “Dynamic arm swinging
in human walking.” In: Proceedings of the Royal Society of London B: Biological Sciences
276.1673 (2009), pp. 3679–3688.

[32] Steven H Collins and Andy Ruina. “A bipedal walking robot with efficient and
human-like gait.” In: Robotics and Automation, 2005. ICRA 2005. Proceedings of the
2005 IEEE International Conference on. IEEE. 2005, pp. 1983–1988.

[33] Hongkai Dai, Andrés Valenzuela, and Russ Tedrake. “Whole-body motion planning
with centroidal dynamics and full kinematics.” In: Humanoid Robots (Humanoids),
2014 14th IEEE-RAS International Conference on. IEEE. 2014, pp. 295–302.

[34] Tomas De Boer. Foot placement in robotic bipedal locomotion. TU Delft, Delft University
of Technology, 2012.

[35] Avik De and Daniel E Koditschek. “The Penn Jerboa: A platform for exploring
parallel composition of templates.” In: arXiv preprint arXiv:1502.05347 (2016).

[36] Tom Erez and Emanuel Todorov. “Trajectory optimization for domains with contacts
using inverse dynamics.” In: Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ
International Conference on. IEEE. 2012, pp. 4914–4919.
 bibliography 82

[37] Michael Ernst, Hartmut Geyer, and Reinhard Blickhan. “Spring-legged locomotion
on uneven ground: a control approach to keep the running speed constant.” In:
International Conference on Climbing and Walking Robots (CLAWAR). World Scientific.
2009, pp. 639–644.

[38] Roy Featherstone. “The acceleration vector of a rigid body.” In: The International
Journal of Robotics Research 20.11 (2001), pp. 841–846.

[39] Roy Featherstone. “Plucker basis vectors.” In: Robotics and Automation, 2006. ICRA
2006. Proceedings 2006 IEEE International Conference on. IEEE. 2006, pp. 1892–1897.

[40] Roy Featherstone. “A beginner’s guide to 6-D vectors (part 2).” In: IEEE robotics &
automation magazine 17.4 (2010), pp. 88–99.

[41] Roy Featherstone. “A beginner’s guide to 6-d vectors (part 1).” In: IEEE robotics &
automation magazine 17.3 (2010), pp. 83–94.

[42] Roy Featherstone. Rigid body dynamics algorithms. Springer, 2014.

[43] Roy Featherstone and David Orin. “Robot dynamics: equations and algorithms.”
In: Robotics and Automation, 2000. Proceedings. ICRA’00. IEEE International Conference
on. Vol. 1. IEEE. 2000, pp. 826–834.

[44] Martin L Felis. “RBDL: an efficient rigid-body dynamics library using recursive
algorithms.” In: Autonomous Robots 41.2 (2017), pp. 495–511.

[45] Paulo Flores, Margarida Machado, Eurico Seabra, and Miguel Tavares da Silva. “A
parametric study on the Baumgarte stabilization method for forward dynamics of
constrained multibody systems.” In: Journal of computational and nonlinear dynamics
6.1 (2011), p. 011019.

[46] Gerrit A Folkertsma, Sangbae Kim, and Stefano Stramigioli. “Parallel stiffness in a
bounding quadruped with flexible spine.” In: Intelligent Robots and Systems (IROS),
2012 IEEE/RSJ International Conference on. IEEE. 2012, pp. 2210–2215.

[47] Robert J Full and Daniel E Koditschek. “Templates and anchors: neuromechanical
hypotheses of legged locomotion on land.” In: Journal of Experimental Biology 202.23
(1999), pp. 3325–3332.
 bibliography 83

[48] Hartmut Geyer, Reinhard Blickhan, and Andre Seyfarth. “Natural dynamics of
spring-like running: Emergence of selfstability.” In: 5th International Conference on
Climbing and Walking Robots. Suffolk, England: Professional Engineering Publishing
Ltd. 2002, pp. 87–91.

[49] Hartmut Geyer, Andre Seyfarth, and Reinhard Blickhan. “Spring-mass running: sim-
ple approximate solution and application to gait stability.” In: Journal of theoretical
biology 232.3 (2005), pp. 315–328.

[50] Hartmut Geyer, Andre Seyfarth, and Reinhard Blickhan. “Compliant leg behaviour
explains basic dynamics of walking and running.” In: Proceedings of the Royal Society
of London B: Biological Sciences 273.1603 (2006), pp. 2861–2867.

[51] David E Goldberg and Kalyanmoy Deb. “A comparative analysis of selection

schemes used in genetic algorithms.” In: Foundations of genetic algorithms. Vol. 1.
Elsevier, 1991, pp. 69–93.

[52] Clement Gosselin and Jorge Angeles. “A global performance index for the kinematic
optimization of robotic manipulators.” In: Journal of Mechanical Design 113.3 (1991),
pp. 220–226.

[53] Jesse A Grimes. “ATRIAS 1.0 & 2.1: enabling agile biped locomotion with a template-
driven approach to robot design.” In: (2013).

[54] Jesse A Grimes and Jonathan W Hurst. “The design of ATRIAS 1.0 a unique
monopod, hopping robot.” In: Adaptive Mobile Robotics. World Scientific, 2012,
pp. 548–554.

[55] Duncan W Haldane, MM Plecnik, Justin K Yim, and Ronald S Fearing. “Robotic
vertical jumping agility via series-elastic power modulation.” In: Science Robotics 1.1
(2016), eaag2048.

[56] R van Ham, Thomas G Sugar, Bram Vanderborght, Kevin W Hollander, and Dirk
Lefeber. “Compliant actuator designs.” In: Robotics & Automation Magazine, IEEE
16.3 (2009), pp. 81–94.
 bibliography 84

[57] Ayonga Hereid, Eric A Cousineau, Christian M Hubicki, and Aaron D Ames. “3D
Dynamic Walking with Underactuated Humanoid Robots: A Direct Collocation
Framework for Optimizing Hybrid Zero Dynamics.” In: IEEE International Conference
on Robotics and Automation (ICRA). IEEE. 2016.

[58] Shigeo Hirose. “Some considerations on a feasible walking mechanism as a terrain

vehicle.” In: 3rd CISM-IFToMM Int. Symp. on Theory and Practice of Robots and
Manipulators. 1978, pp. 357–375.

[59] Philip Holmes, Robert J Full, Dan Koditschek, and John Guckenheimer. “The
dynamics of legged locomotion: Models, analyses, and challenges.” In: Siam Review
48.2 (2006), pp. 207–304.

[60] Christian M Hubicki and Jonathan W Hurst. “Running on soft ground: Simple,
energy-optimal disturbance rejection.” In: International Conference on Climbing and
Walking Robots (CLAWAR). 2012.

[61] Christian Michael Hubicki. “From running birds to walking robots: optimization as
a unifying framework for dynamic bipedal locomotion.” In: (2014).

[62] Christian Hubicki, Andy Abate, Patrick Clary, Siavash Rezazadeh, Mikhail Jones,
Andrew Peekema, Johnathan Van Why, Ryan Domres, Albert Wu, William Martin,
et al. “Walking and running with passive compliance: Lessons from engineering a
live demonstration of the atrias biped.” In: IEEE Robotics and Automation Magazine
().

[63] Christian Hubicki, Jesse Grimes, Mikhail Jones, Daniel Renjewski, Alexander
Spröwitz, Andy Abate, and Jonathan Hurst. “ATRIAS: Design and validation
of a tether-free 3D-capable spring-mass bipedal robot.” In: The International Journal
of Robotics Research 35.12 (2016), pp. 1497–1521.

[64] Jonathan W Hurst. “The role and implementation of compliance in legged locomo-
tion.” PhD thesis. Carnegie Mellon University, 2008.
 bibliography 85

[65] Marco Hutter, C David Remy, Mark A Hoepflinger, and Roland Siegwart. “Scarleth:
Design and control of a planar running robot.” In: Intelligent Robots and Systems
(IROS), 2011 IEEE/RSJ International Conference on. IEEE. 2011, pp. 562–567.

[66] Marco Hutter, Michael Gehring, Michael Bloesch, C David Remy, Roland Yves
Siegwart, and Mark A Hoepflinger. “StarlETH: A compliant quadrupedal robot for
fast, efficient, and versatile locomotion.” In: (2012).

[67] Sang-Ho Hyon and Tsutomu Mita. “Development of a biologically inspired hopping
robot-" Kenken".” In: Robotics and Automation, 2002. Proceedings. ICRA’02. IEEE
International Conference on. Vol. 4. IEEE. 2002, pp. 3984–3991.

[68] Mikhail S Jones. “Optimal control of an underactuated bipedal robot.” In: (2014).

[69] Mikhail S Jones and Jonathan W Hurst. “Effects of leg configuration on running
and walking robots.” In: Proceedings of the 5th international conference on climbing
and walking robots and the support technologies for mobile machines, Baltimore. 2012,
pp. 519–526.

[70] Anders Jönsson, J Bathelt, and G Broman. “Implications of modelling one-dimensional

impact by using a spring and damper element.” In: Proceedings of the Institution of
Mechanical Engineers, Part K: Journal of Multi-body Dynamics 219.3 (2005), pp. 299–305.

[71] Shuuji Kajita, Fumio Kanehiro, Kenji Kaneko, Kiyoshi Fujiwara, Kensuke Harada,
Kazuhito Yokoi, and Hirohisa Hirukawa. “Resolved momentum control: Humanoid
motion planning based on the linear and angular momentum.” In: Intelligent Robots
and Systems, 2003.(IROS 2003). Proceedings. 2003 IEEE/RSJ International Conference on.
Vol. 2. IEEE. 2003, pp. 1644–1650.

[72] Carl T Kelley. Iterative methods for optimization. SIAM, 1999.

[73] Gavin Kenneally, Avik De, and Daniel E Koditschek. “Design principles for a family
of direct-drive legged robots.” In: IEEE Robotics and Automation Letters 1.2 (2016),
pp. 900–907.

[74] Ben Kenwright. “Inverse kinematics with dual-quaternions, exponential-maps, and

joint limits.” In: (2013).
 bibliography 86

[75] Oussama Khatib. “Real-time obstacle avoidance for manipulators and mobile
robots.” In: Autonomous robot vehicles. Springer, 1986, pp. 396–404.

[76] Oussama Khatib. “Inertial properties in robotic manipulation: An object-level

framework.” In: The international journal of robotics research 14.1 (1995), pp. 19–36.

[77] Sangbae Kim and Patrick M. Wensing. “Design of Dynamic Legged Robots.” In:
Foundations and Trends® in Robotics 5.2 (2017), pp. 117–190. doi: 10.1561/2300000044.
url: http://dx.doi.org/10.1561/2300000044.

[78] Devin Koepl and Jonathan Hurst. “Force control for planar spring-mass running.”
In: Intelligent Robots and Systems (IROS), 2011 IEEE/RSJ International Conference on.
IEEE. 2011, pp. 3758–3763.

[79] Devin Koepl and Jonathan Hurst. “Impulse control for planar spring-mass run-
ning.” In: Journal of Intelligent & Robotic Systems 74.3-4 (2014), pp. 589–603.

[80] Evangelos Kokkevis. “Practical physics for articulated characters.” In: Game Devel-
opers Conference. Vol. 2004. 2004.

[81] Arthur D Kuo. “Choosing your steps carefully.” In: IEEE Robotics & Automation
Magazine 14.2 (2007), pp. 18–29.

[82] Michael LaBarbera. “Why the wheels won’t go.” In: The American Naturalist 121.3
(1983), pp. 395–408.

[83] Dominic Lakatos, Werner Friedl, and Alin Albu-Schäffer. “Eigenmodes of Nonlinear
Dynamics: Definition, Existence, and Embodiment into Legged Robots With Elastic
Elements.” In: IEEE Robotics and Automation Letters 2.2 (2017), pp. 1062–1069.

[84] Jimmy Kwok-Ching Lam. “An efficient simulated annealing schedule.” PhD thesis.
Yale University, 1989.

[85] David V Lee and Sanford G Meek. “Directionally compliant legs influence the
intrinsic pitch behaviour of a trotting quadruped.” In: Proceedings of the Royal Society
of London B: Biological Sciences 272.1563 (2005), pp. 567–572.
 bibliography 87

[86] Jihong Lee. “A study on the manipulability measures for robot manipulators.”
In: Intelligent Robots and Systems, 1997. IROS’97., Proceedings of the 1997 IEEE/RSJ
International Conference on. Vol. 3. IEEE. 1997, pp. 1458–1465.

[87] Sigrid Leyendecker. “Mechanical integrators for constrained dynamical systems in

flexible multibody dynamics.” PhD thesis. 2006.

[88] Zexiang Li and Richard Montgomery. “Dynamics and optimal control of a legged
robot in flight phase.” In: Robotics and Automation, 1990. Proceedings., 1990 IEEE
International Conference on. IEEE. 1990, pp. 1816–1821.

[89] Zhibin Li, Nikos G Tsagarakis, and Darwin G Caldwell. “A passivity based ad-
mittance control for stabilizing the compliant humanoid COMAN.” In: Humanoid
Robots (Humanoids), 2012 12th IEEE-RAS International Conference on. IEEE. 2012,
pp. 43–49.

[90] Imed Mansouri and Mohammed Ouali. “The power manipulability–A new ho-
mogeneous performance index of robot manipulators.” In: Robotics and Computer-
Integrated Manufacturing 27.2 (2011), pp. 434–449.

[91] William C Martin, Albert Wu, and Hartmut Geyer. “Robust spring mass model
running for a physical bipedal robot.” In: Robotics and Automation (ICRA), 2015 IEEE
International Conference on. IEEE. 2015, pp. 6307–6312.

[92] H-M Maus, SW Lipfert, M Gross, J Rummel, and A Seyfarth. “Upright human
gait did not provide a major mechanical challenge for our ancestors.” In: Nature
communications 1 (2010), p. 70.

[93] Tad McGeer. “Powered flight, child’s play, silly wheels and walking machines.”
In: Robotics and Automation, 1989. Proceedings., 1989 IEEE International Conference on.
IEEE. 1989, pp. 1592–1597.

[94] Tad McGeer et al. “Passive dynamic walking.” In: I. J. Robotic Res. 9.2 (1990), pp. 62–
82.
 bibliography 88

[95] Robert B McGhee and Geoffrey I Iswandhi. “Adaptive locomotion of a multilegged

robot over rough terrain.” In: IEEE transactions on systems, man, and cybernetics 9.4
(1979), pp. 176–182.

[96] Andreas Merker, Dieter Kaiser, and Martin Hermann. “Numerical bifurcation
analysis of the bipedal spring-mass model.” In: Physica D: Nonlinear Phenomena 291
(2015), pp. 21–30.

[97] Jean-Pierre Merlet. “Jacobian, manipulability, condition number, and accuracy of

parallel robots.” In: Journal of Mechanical Design 128.1 (2006), pp. 199–206.

[98] Igor Mordatch. “Automated Discovery and Learning of Complex Movement Behav-
iors.” PhD thesis. 2016.

[99] Igor Mordatch, Emanuel Todorov, and Zoran Popović. “Discovery of complex
behaviors through contact-invariant optimization.” In: ACM Transactions on Graphics
(TOG) 31.4 (2012), p. 43.

[100] Xiaohong Nian, Fei Peng, and Hang Zhang. “Regenerative braking system of
electric vehicle driven by brushless DC motor.” In: IEEE Transactions on Industrial
Electronics 61.10 (2014), pp. 5798–5808.

[101] Ryuma Niiyama and Yasuo Kuniyoshi. “Design principle based on maximum
output force profile for a musculoskeletal robot.” In: Industrial Robot: An International
Journal 37.3 (2010), pp. 250–255.

[102] Ionut Mihai Constantin Olaru, Chris Schmidt-Wetekam, Nicholas Payton, Gray
Thomas, Johnny Godowski, Jerry Pratt, and Sebastien Cotton. “Mechanical Design
of the FastRunner Leg.” In: HIP 10.90 (), p. 100.

[103] Kevin C Olds. “Global Indices for kinematic and force transmission performance in
parallel robots.” In: Robotics, IEEE Transactions on 31.2 (2015), pp. 494–500.

[104] David E Orin and Ambarish Goswami. “Centroidal momentum matrix of a hu-
manoid robot: Structure and properties.” In: Intelligent Robots and Systems, 2008.
IROS 2008. IEEE/RSJ International Conference on. IEEE. 2008, pp. 653–659.
 bibliography 89

[105] David E Orin, Ambarish Goswami, and Sung-Hee Lee. “Centroidal dynamics of a
humanoid robot.” In: Autonomous Robots 35.2-3 (2013), pp. 161–176.

[106] Toru Oshima, Tomohiko Fujikawa, Osamu Kameyama, and Minayori Kumamoto.
“Robotic analyses of output force distribution developed by human limbs.” In: Robot
and Human Interactive Communication, 2000. RO-MAN 2000. Proceedings. 9th IEEE
International Workshop on. IEEE. 2000, pp. 229–234.

[107] F Ozguner, SJ Tsai, and RB McGhee. “An approach to the use of terrain-preview
information in rough-terrain locomotion by a hexapod walking machine.” In: The
International Journal of Robotics Research 3.2 (1984), pp. 134–146.

[108] Jim M Papadopoulos. “Forces in bicycle pedalling.” In: Biomechanics in Sport (1987),
pp. 26–33.

[109] KV Papantoniou. “Electromechanical design for an electrically powered, actively

balanced one leg planar robot.” In: Intelligent Robots and Systems’ 91.’Intelligence for
Mechanical Systems, Proceedings IROS’91. IEEE/RSJ International Workshop on. IEEE.
1991, pp. 1553–1560.

[110] Frank C Park, James E Bobrow, and Scott R Ploen. “A Lie group formulation
of robot dynamics.” In: The International Journal of Robotics Research 14.6 (1995),
pp. 609–618.

[111] Hae-Won Park, Koushil Sreenath, Jonathan W Hurst, and Jessy W Grizzle. “Identi-
fication of a bipedal robot with a compliant drivetrain.” In: IEEE Control Systems
Magazine 31.2 (2011), pp. 63–88.

[112] Frank Peuker, Christophe Maufroy, and André Seyfarth. “Leg-adjustment strategies
for stable running in three dimensions.” In: Bioinspiration & biomimetics 7.3 (2012),
p. 036002.

[113] Michael Posa. “Optimization for Control and Planning of Multi-contact Dynamic
Motion.” PhD thesis. Massachusetts Institute of Technology, 2017.
 bibliography 90

[114] Gill A Pratt and Matthew M Williamson. “Series elastic actuators.” In: Intelligent
Robots and Systems 95.’Human Robot Interaction and Cooperative Robots’, Proceedings.
1995 IEEE/RSJ International Conference on. Vol. 1. IEEE. 1995, pp. 399–406.

[115] Jerry E Pratt and Gill A Pratt. “Exploiting natural dynamics in the control of a
planar bipedal walking robot.” In: Proceedings of the Annual Allerton Conference on
Communication Control and Computing. Vol. 36. UNIVERSITY OF ILLINOIS. 1998,
pp. 739–748.

[116] Jerry Pratt, John Carff, Sergey Drakunov, and Ambarish Goswami. “Capture point:
A step toward humanoid push recovery.” In: Humanoid Robots, 2006 6th IEEE-RAS
International Conference on. IEEE. 2006, pp. 200–207.

[117] Marc H Raibert, H Benjamin Brown, and Michael Chepponis. “Experiments in

balance with a 3D one-legged hopping machine.” In: The International Journal of
Robotics Research 3.2 (1984), pp. 75–92.

[118] Marc H Raibert, H Benjamin Brown Jr, Michael Chepponis, Jeff Koechling, Jessica K
Hodgins, Diane Dustman, W Kevin Brennan, David S Barrett, Clay M Thompson,
John Daniell Hebert, et al. “Dynamically Stable Legged Locomotion (September
1985-Septembers1989).” In: (1989).

[119] Marc Raibert, Kevin Blankespoor, Gabriel Nelson, and Rob Playter. “Bigdog, the
rough-terrain quadruped robot.” In: IFAC Proceedings Volumes 41.2 (2008), pp. 10822–
10825.

[120] Jacob Reher, Eric A Cousineau, Ayonga Hereid, Christian M Hubicki, and Aaron D
Ames. “Realizing dynamic and efficient bipedal locomotion on the humanoid robot
DURUS.” In: Robotics and Automation (ICRA), 2016 IEEE International Conference on.
IEEE. 2016, pp. 1794–1801.

[121] C David Remy, Keith Buffinton, and Roland Siegwart. “A matlab framework for
efficient gait creation.” In: Intelligent Robots and Systems (IROS), 2011 IEEE/RSJ
International Conference on. IEEE. 2011, pp. 190–196.
 bibliography 91

[122] Siavash Rezazadeh and Jonathan W Hurst. “On the optimal selection of motors and
transmissions for electromechanical and robotic systems.” In: Intelligent Robots and
Systems (IROS 2014), 2014 IEEE/RSJ International Conference on. IEEE. 2014, pp. 4605–
4611.

[123] Siavash Rezazadeh and Jonathan W Hurst. “Toward step-by-step synthesis of

stable gaits for underactuated compliant legged robots.” In: Robotics and Automation
(ICRA), 2015 IEEE International Conference on. IEEE. 2015, pp. 4532–4538.

[124] Siavash Rezazadeh, Christian Hubicki, Mikhail Jones, Andrew Peekema, Johnathan
Van Why, Andy Abate, and Jonathan Hurst. “Spring-mass walking with ATRIAS
in 3D: Robust gait control spanning zero to 4.3 kph on a heavily underactuated
bipedal robot.” In: ASME 2015 Dynamic Systems and Control Conference. American
Society of Mechanical Engineers. 2015, V001T04A003–V001T04A003.

[125] Thomas J Roberts and Emanuel Azizi. “Flexible mechanisms: the diverse roles of
biological springs in vertebrate movement.” In: The Journal of experimental biology
214.3 (2011), pp. 353–361.

[126] David W Robinson, Jerry E Pratt, Daniel J Paluska, and Gill A Pratt. “Series elastic
actuator development for a biomimetic walking robot.” In: Advanced Intelligent
Mechatronics, 1999. Proceedings. 1999 IEEE/ASME International Conference on. IEEE.
1999, pp. 561–568.

[127] Fredrik Roos, Hans Johansson, and Jan Wikander. “Optimal selection of motor and
gearhead in mechatronic applications.” In: Mechatronics 16.1 (2006), pp. 63–72.

[128] Martin Rutschmann, Brian Satzinger, Marten Byl, and Katie Byl. “Nonlinear model
predictive control for rough-terrain robot hopping.” In: Intelligent Robots and Systems
(IROS), 2012 IEEE/RSJ International Conference on. IEEE. 2012, pp. 1859–1864.

[129] Uluc Saranli, Martin Buehler, and Daniel E Koditschek. “RHex: A simple and highly
mobile hexapod robot.” In: The International Journal of Robotics Research 20.7 (2001),
pp. 616–631.
 bibliography 92

[130] John Schmitt, Mariano Garcia, RC Razo, Philip Holmes, and Robert J Full. “Dy-
namics and stability of legged locomotion in the horizontal plane: a test case using
insects.” In: Biological cybernetics 86.5 (2002), pp. 343–353.

[131] William J Schwind and Daniel E Koditschek. “Approximating the stance map of a
2-DOF monoped runner.” In: Journal of Nonlinear Science 10.5 (2000), pp. 533–568.

[132] Claudio Semini, Victor Barasuol, Thiago Boaventura, Marco Frigerio, Michele
Focchi, Darwin G Caldwell, and Jonas Buchli. “Towards versatile legged robots
through active impedance control.” In: The International Journal of Robotics Research
34.7 (2015), pp. 1003–1020.

[133] Sangok Seok, Albert Wang, Meng Yee Michael Chuah, Dong Jin Hyun, Jongwoo
Lee, David M Otten, Jeffrey H Lang, and Sangbae Kim. “Design principles for
energy-efficient legged locomotion and implementation on the MIT Cheetah robot.”
In: IEEE/ASME Transactions on Mechatronics 20.3 (2015), pp. 1117–1129.

[134] André Seyfarth, Hartmut Geyer, and Hugh Herr. “Swing-leg retraction: a simple
control model for stable running.” In: Journal of Experimental Biology 206.15 (2003),
pp. 2547–2555.

[135] Zhuohua Shen and Justin Seipel. “A spring-mass model of locomotion with full
asymptotic stability.” In: ASME 2011 International Mechanical Engineering Congress
and Exposition. American Society of Mechanical Engineers. 2011, pp. 301–306.

[136] Zhuohua Shen and Justin Seipel. “A Piecewise-Linear Approximation of the Canon-
ical Spring-Loaded Inverted Pendulum Model of Legged Locomotion.” In: Journal
of Computational and Nonlinear Dynamics 11.1 (2016), p. 011007.

[137] Manuel F Silva and JA Tenreiro Machado. “A historical perspective of legged

robots.” In: Journal of Vibration and Control 13.9-10 (2007), pp. 1447–1486.

[138] Peng Song. “Modeling, analysis and simulation of multibody systems with contact
and friction.” PhD thesis. University of Pennsylvania, 2002.

[139] Shin-Min Song and Jong-Kil Lee. “The mechanical efficiency and kinematics of
pantograph-type manipulators.” In: KSME Journal 2.1 (1988), pp. 69–78.
 bibliography 93

[140] James C Spall. Introduction to stochastic search and optimization: estimation, simulation,
and control. Vol. 65. John Wiley & Sons, 2005.

[141] Manoj Srinivasan and Andy Ruina. “Computer optimization of a minimal biped
model discovers walking and running.” In: Nature 439.7072 (2006), pp. 72–75.

[142] Hai-Jun Su. “A pseudorigid-body 3R model for determining large deflection of

cantilever beams subject to tip loads.” In: Journal of Mechanisms and Robotics 1.2
(2009), p. 021008.

[143] NA Titus and CH Spenny. “Power metrics for robot planning and redundancy
resolution.” In: Intelligent Control, 1994., Proceedings of the 1994 IEEE International
Symposium on. IEEE. 1994, pp. 153–159.

[144] Emanuel Todorov. “Convex and analytically-invertible dynamics with contacts and
constraints: Theory and implementation in MuJoCo.” In: Robotics and Automation
(ICRA), 2014 IEEE International Conference on. IEEE. 2014, pp. 6054–6061.

[145] Emanuel Todorov, Tom Erez, and Yuval Tassa. “Mujoco: A physics engine for model-
based control.” In: Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International
Conference on. IEEE. 2012, pp. 5026–5033.

[146] Nikos G Tsagarakis, Stephen Morfey, Gustavo Medrano Cerda, Li Zhibin, and
Darwin G Caldwell. “Compliant humanoid coman: Optimal joint stiffness tuning for
modal frequency control.” In: Robotics and Automation (ICRA), 2013 IEEE International
Conference on. IEEE. 2013, pp. 673–678.

[147] James R Usherwood and John EA Bertram. “Understanding brachiation: insight

from a collisional perspective.” In: Journal of Experimental Biology 206.10 (2003),
pp. 1631–1642.

[148] Hamid Reza Vejdani Noghreiyan. “Control of spring-mass running robots.” In:
(2013).

[149] Hamid Reza Vejdani, Albert Wu, Hartmut Geyer, and Jonathan W Hurst. “Touch-
down angle control for spring-mass walking.” In: Robotics and Automation (ICRA),
2015 IEEE International Conference on. IEEE. 2015, pp. 5101–5106.
 bibliography 94

[150] Miomir Vukobratović and Branislav Borovac. “Zero-moment point—thirty five

years of its life.” In: International journal of humanoid robotics 1.01 (2004), pp. 157–173.

[151] KJ Waldron and GL Kinzel. “The relationship between actuator geometry and
mechanical efficiency in robots.” In: Fourth, symposium on Theory and Practice of
Robots and Manipulators. Poland (1981).

[152] Handing Wang and Xin Yao. “Corner sort for Pareto-based many-objective opti-
mization.” In: IEEE transactions on cybernetics 44.1 (2014), pp. 92–102.

[153] Patrick M Wensing. “Optimization and control of dynamic humanoid running and
jumping.” PhD thesis. 2014.

[154] Patrick M Wensing, Albert Wang, Sangok Seok, David Otten, Jeffrey Lang, and
Sangbae Kim. “Proprioceptive actuator design in the MIT cheetah: Impact mitigation
and high-bandwidth physical interaction for dynamic legged robots.” In: IEEE
Transactions on Robotics (2017).

[155] Patrick Wensing, Roy Featherstone, and David E Orin. “A reduced-order recursive
algorithm for the computation of the operational-space inertia matrix.” In: Robotics
and Automation (ICRA), 2012 IEEE International Conference on. IEEE. 2012, pp. 4911–
4917.

[156] Yonas Gebre Woldesenbet, Gary G Yen, and Biruk G Tessema. “Constraint handling
in multiobjective evolutionary optimization.” In: IEEE Transactions on Evolutionary
Computation 13.3 (2009), pp. 514–525.

[157] Aimin Wu and Hartmut Geyer. “The 3-d spring–mass model reveals a time-based
deadbeat control for highly robust running and steering in uncertain environments.”
In: Robotics, IEEE Transactions on 29.5 (2013), pp. 1114–1124.

[158] Ming-Ji Yang, Hong-Lin Jhou, Bin-Yen Ma, and Kuo-Kai Shyu. “A cost-effective
method of electric brake with energy regeneration for electric vehicles.” In: IEEE
Transactions on Industrial Electronics 56.6 (2009), pp. 2203–2212.

[159] Tsuneo Yoshikawa. “Manipulability of robotic mechanisms.” In: The international

journal of Robotics Research 4.2 (1985), pp. 3–9.
 bibliography 95

[160] Petr Zaytsev, S Javad Hasaneini, and Andy Ruina. “Two steps is enough: no need
to plan far ahead for walking balance.” In: Robotics and Automation (ICRA), 2015
IEEE International Conference on. IEEE. 2015, pp. 6295–6300.

[161] Garth John Zeglin. “Uniroo–a one legged dynamic hopping robot.” PhD thesis.
Massachusetts Institute of Technology, 1991.
colophon

This document was typeset using the typographical look-and-feel classicthesis devel-
oped by André Miede. The style was inspired by Robert Bringhurst’s seminal book on
typography “The Elements of Typographic Style”. classicthesis is available for both LATEX
and LYX:

http://code.google.com/p/classicthesis/

Final Version as of May 22, 2018 (classicthesis version 4.2).