INTEG. AND COMP. BIOL.

, 42:149157 (2002)

Quantifying Dynamic Stability and Maneuverability in Legged Locomotion1

ROBERT J. FULL,2,* TIMOTHY KUBOW,* JOHN SCHMITT, PHILIP HOLMES, AND DANIEL KODITSCHEK
*Department of Integrative Biology, University of California, Berkeley, California 94720
Department of Mechanical and Aerospace Engineering, Princeton University, Princeton, New Jersey 80544
Department of Electrical Engineering and Computer Science, University of Michigan,
Ann Arbor, Michigan 48109-2110

SYNOPSIS. Animals can swerve, dodge, dive, climb, turn and stop abruptly. Their stability and maneuver-
ability are remarkable, but a challenge to quantify. Formal stability analysis can allow for quantitative
comparisons within and among species. Stability analysis used in concert with a template (a simple, general
model that serves as a guide for control) can lead to testable hypotheses of function. Neural control models
postulated without knowledge of the animals mechanical (musculo-skeletal) system can be counterproduc-
tive and even destabilizing. Perturbations actively corrected by reflex feedback in one direction can result
in perturbations in other directions because the system is coupled dynamically. The passive rate of recovery
from a perturbation in one direction differs from rates in other directions. We hypothesize that animals
might exert less neural control in directions that rapidly recover via passive dynamics (e.g., in body orien-
tation and rotation). By contrast, animals are likely to exert more neural control in directions that recover
slowly or not at all via passive dynamics (e.g., forward velocity and heading). Neural control best enhances
stability when it works with the natural, passive dynamics of the mechanical system. Measuring maneuver-
ability is more challenging and new, general metrics are needed. Templates reveal that simple analyses of
summed forces and quantification of the center of pressure can lead to valuable hypotheses, whereas kine-
matic descriptions may be inadequate. The study of stability and maneuverability has direct relevance to
the behavior and ecology of animals, but is also critical if animal design is to be understood. Animals appear
to be grossly over-built for steady-state, straight-ahead locomotion, as they appear to possess too many
neurons, muscles, joints and even too many appendages. The next step in animal locomotion is to subject
animals to perturbations and reveal the function of all their parts.

INTRODUCTION techniques and the common language required in these

Animals can negotiate the most complex terrain, sub-disciplines, these communities have been some-
turn in an instant and stop on a dime. We contend that what isolated from one another. The time has come for
the time has come to propose simple, testable, quan- an integrative view of locomotion (Dickinson et al.,
titative hypotheses explaining this remarkable ability. 2000). Third, the study of stability and maneuverabil-
The mechanisms that allow extraordinary stability and ity requires a truly interdisciplinary effort. Relevant,
maneuverability have remained largely unexplored for general control models have been developed in fields
several reasons. First, the methodology and search for as diverse as comparative biology (Cruse, 1990; Duy-
simple, general principles initially demanded the study sens et al., 2000; Pearson, 1993; McIntyre and Bizzi,
of steady-state, straight-ahead locomotion. To charac- 1993), the human biomechanics of arm movements
terize more complex responses to perturbations, one (Hogan, 1990), walking and maintaining posture (Bau-
must first define the steady-state condition to which by and Kuo, 2000; Mackinnon and Winter, 1993; Zehr
the animal will return. Just as the fields of exercise and Stein, 1999), and in the motor control branch of
physiology, gas transport and energy utilization first psychology focusing on hand movements (Hanken et
advanced our understanding of the steady-state (e.g., al., 1985). As is shown by the recent approaches in
Taylor and Weibel, 1981; Taylor et al., 1982) and are these diverse areas, we argue that a principled under-
now taking the next logical step toward the study of standing will not only draw together fields within bi-
intermittent locomotion (see Amer. Zool., 41:137246, ology broadly construed, but must integrate biology
2001), so should locomotor biomechanics and neural with engineering and mathematics.
control move beyond the steady-state. Second, suffi- In the present publication we define stability, briefly
cient knowledge of individual systems in traditional introduce an established method to quantify stability,
disciplines is required before their interactions can be show how quantifying the stability of a dynamic model
explored most productively. Useful hypotheses are can lead to novel hypotheses of control and animal
available for individual neurons, muscles, feedforward design, and discuss simple measurements of maneu-
pattern generators and feedback circuits functioning in verability. Although we analyze legged locomotion in
moving animals. Due to the focused efforts, shared particular, the approach can be used for all modes of
locomotion and behavior in general. We hope to en-
1 From the Symposium Stability and Maneuverability presented at
courage investigators, particularly those who collect
the Annual Meeting of the Society for Integrative and Comparative
kinematic data, to consider the value of dynamical sta-
Biology, 37 January 2001, at Chicago, Illinois. bility and maneuverability analyses. Our primary goal
2 E-mail: [email protected] is to further articulate a paradigm for understanding

149
150 R. J. FULL ET AL.

the mechanisms of locomotion by using these analyses ability of characteristic measurements (i.e., state vari-
to formulate and test hypotheses generated from dy- ables such as velocities, angles and positions) to return
namic models (Full and Koditschek, 1999). We con- to a steady-state, periodic gait after a perturbation.
tend that at least two interconnected classes of dynam-
ic models are required to integrate function across the Quantifying dynamic stabilitydynamical systems
various levels of organization. One model that we term theory
a template, incorporates the fewest possible degrees The field of dynamical systems provides an estab-
of freedom and serves as a control target (Full and lished methodology to quantify stability. Our aim in
Koditschek, 1999). The second model, the anchor, the present publication is not to explain the details of
is a more complex physically realistic model within dynamical systems theory (see Guckenheimer and
which the template behavior is embedded. We argue Holmes, 1983; Strogatz, 1994), but to give sufficient
that efforts to build templates and to anchor them in background so that those studying locomotion can see
realistic animal data represent a productive avenue to- its potential in description and hypothesis formation.
ward an integrated view of neuromechanical control. It is important to note that dynamical systems theory
Defining stability is a challenge involves the formal analysis of how systems at any
level of organization (neuron, networks or behaviors)
Static stability. Traditionally, stability in legged lo- change over time. In this context, the term dynamical
comotion is taken to refer to static stability. The ne- system is not restricted to a system generating forces
cessity for static stability in arthropods has been used (kinetics) and moving (kinematics), as is the common
as one of, if not the most important, reason, why in- usage in biomechanics. The description of stability re-
sects have at least six legs and use two sets of alter- sulting from dynamical systems theory, which address-
nating tripods of support during locomotion (Hughes, es mathematical models, differential equations and it-
1952). Numerous investigators (Alexander, 1982; Del-
erated mappings, does not necessarily provide us with
comyn, 1985; Graham, 1985; Hughes, 1952) have dis-
a direct correspondence to a particular biomechanical
cussed the stepping patterns that insects require to
structure (Beck et al., 1995). Instead, the resulting sta-
maintain static stability during locomotion. Yet, few
bility analysis acts to guide our attention in productive
have attempted to quantify static stability as a function
directions to search for just such a link between co-
of gait or variation in body form (Cruse, 1976;
ordination hypotheses from dynamical systems and
Hughes, 1957). Research on legged walking machines
mechanisms based in biomechanics and motor control
provided an approach to quantify static stability
(e.g., Taga, 1995a, b; Taga et al., 1991; Dingwell and
(McGhee, 1976; Song, 1984; Sun, 1974). The mini-
Cusumano, 2000).
mum requirement to attain static stability is a tripod
Define and measure variables that specify the state
of support, as in a stool. If an animals center of mass
falls outside the triangle of support formed by its three of the system. The first task in the quantification of
feet on the ground, it is statically unstable and will stability is to decide on what is best to measure. The
fall. In the quasi-static gait of a robot or animal, the goal is to specify a set of variables such as positions
center of mass moves with respect to the legs, and the and velocities that completely define the state of the
likelihood of falling increases the closer the center of system. State variables are distinct from parameters
mass comes to the edge of the triangle of support. such as mass, inertia and leg length that are more or
Dynamic stability analysis is required for all but the less fixed for a given animal. State variables change
slowest movements. Ting et al. (1994) discovered that over time as determined by the dynamics of the sys-
the degree of static stability decreased as insects ran tem. Ideally, their values at any instant in time should
faster, until at the highest speeds they became statically allow the determination of all future values. Put an-
unstable during certain parts of each stride, even when other way, if two different trials of a running animal
a support tripod was present. In the sagittal plane, six converge to the same values, their locomotion patterns
and eight legged animals are best modeled as dynamic, should be very similar from that time forward.
spring-load, inverted pendula (Blickhan and Full, Periodic trajectories called limit cycles characterize
1987; Full and Tu, 1990) in the same way as two- and locomotion. During stable, steady-state locomotion the
four-legged runners (Blickhan and Full, 1993; Cav- value of state variables oscillates rhythmically over
agna et al., 1977). At the highest speeds, ghost crabs, time (e.g., lateral velocity in Fig. 1A). In addition to
cockroaches and ants exhibit aerial phases (Blickhan representing the behavior of the state variables with
and Full, 1987; Full and Tu, 1991; Zollikofer, 1994). respect to time, we can examine their behavior relative
In the horizontal plane, insects and other legged run- to one another. Figure 1B shows a plot of the state
ners are best modeled by a dynamic, lateral leg spring, variables (e.g., lateral, rotational and fore-aft velocity)
bouncing the animal from side to side (Schmitt and in state space. Time is no longer an axis, but changes
Holmes, 2000a, b). as one moves along the loop in this three dimensional
These models, and force and velocity measurements space. The closed loop trajectory tells us that the sys-
on animals, suggest that running at a constant average tem is periodic in time. Such a trajectory in state space
speed, while clearly a dynamical process, is essentially is known as a limit cycle. If any other path converges
periodic in time. We define locomotor stability as the to this cycle, it has stabilized to the same trajectory.
 QUANTIFYING DYNAMIC STABILITY 151

FIG. 2. Types of stability. Schematic representations of asymptotic

stability with an equilibrium point (star), neutral stability with a
continuum of equilibrium points, and an example of instability. The
axes represent any two state variables.

state variable is affected as recovery begins, then the

system is said to be uncoupled. In this case, the ei-
genvectors lie along the axes of the state variables.
FIG. 1. Periodic orbit or limit cycle. A. Variation in a single state More often, perturbations delivered in one direction
variable, lateral velocity over one stride. A cycle is present within necessarily result in perturbations in other directions
which lateral velocity repeats from t to t 11. B. Periodic orbit show- because the system is coupled dynamically (Fig. 3D).
ing a limit cycle in state space. Lateral, rotation and fore-aft velocity In a coupled system, the eigenvectors do not coincide
oscillate following a regular trajectory over a stride. Any point in
the cycle can be considered an equilibrium point (star) of the asso-
with the state variables axes. The return path depends
ciated return map. on the relative rates of recovery along the eigenvec-
tors. The rate of recovery from a perturbation in one
direction can differ from rates in other directions. For
Two types of stability existasymptotic and neutral. example, a perturbation increasing lateral velocity
Characterizing stability requires perturbations to state alone could result in a rapid increase in rotational ve-
variables (Fig. 2). Most generally, stability can be de- locity as lateral velocity begins to recover (Fig. 3F).
fined as the ability of a system to return to a stable After several cycles, both rotational and lateral veloc-
limit cycle or equilibrium point after a perturbation. ity may recover back to equilibrium together slowly.
There are at least two types of stable systems. In an The coupled dynamics could be explained by two in-
asymptotically stable system, the return after the per- dependent, orthogonal eigenvectors that differ in rates
turbation is to the original equilibrium or limit cycle. of recovery. The eigenvector showing rapid recovery
In a neutrally stable system, the return to stability after (eigenvalue 5 0.37) pulls the system toward a decrease
perturbation is to a new, nearby, equilibrium or limit in lateral velocity, while increasing rotational velocity
cycle. In an unstable system perturbations tend to (Fig. 3I). The eigenvector showing slow recovery (ei-
grow. genvalue 5 0.72) pulls the system toward a decrease
Quantifying the rate of recovery from a perturba- in both lateral and rotational velocity (Fig. 3L). Ini-
tion. Return of a perturbed state variable to a limit tially, the fast eigenvector dominates, producing partial
cycle can be mapped discretely from cycle to cycle at recovery from the lateral velocity perturbation, but
the same phase (Fig. 3A, B). Mapping all possible per- creating a natural perturbation in rotational velocity.
turbations from one cycle to the next produces a return Later, the slower eigenvector dominates, producing re-
map. A linearized version of the mapping function (an covery in both state variables.
n 3 n matrix, if there are n state variables) can yield Quantifying neutral stability and instability. Stabil-
quantitative measures of stability for a local region ity values (eigenvalues) that approach or equal one im-
(Guckenheimer and Holmes, 1983; Strogatz, 1994). ply that state variables recover slowly to the original
The percentage of the perturbation remaining after equilibrium or not at all. A value of one indicates re-
each cycle is given by the eigenvalue. Eigenvalues covery to a new equilibrium and neutral stability (Fig.
closer to zero indicate a more rapid rate of recovery, 3MO). If the stability value (eigenvalue) is greater
whereas values closer to one indicate a slower rate of than one, then perturbations along that eigenvector
recovery. A stability value (eigenvalue) of 0.37 means grow and the system is unstable (Fig. 3PR).
that only 37% of the perturbation remains after a single
cycle (Fig. 3B). In the next cycle, the remaining per- Forming hypotheses of controlthe importance of
turbation is reduced again by 37%, so that only 13.7% working with the natural dynamics
is left. Reducing the perturbation by a constant per- A formal dynamical analysis can be a powerful
centage after each cycle results in asymptotic recovery. mathematical tool when used for hypothesis genera-
Quantifying the direction of recovery from a per- tion, rather than simply as one of many curve-fitting
turbation. The local direction of recovery is given by techniques. Hypotheses of neural control at the level
the eigenvectors of the linearized mapping function of the whole animal can be generated from the stability
(Fig. 3C). If a state variable is perturbed, and no other analysis of dynamic models. One such template is the
 152
R. J. FULL ET AL.

FIG. 3. Quantifying dynamic stability. First row simulates the change in lateral velocity of the body over time. Second row shows the rate of recovery (eigenvalue) in lateral velocity
of the body over time. The dashed horizontal line in the second row represents the equilibrium value. Third row shows the direction of recovery (eigenvector) in lateral and rotational
velocity of the body. Arrows show the direction of recovery, except in the case of instability. Stars represent equilibrium points. The initial perturbations are represented by open circles.
AC. Asymptotic recovery in an uncoupled systemlateral velocity perturbation. In this model, the perturbation in lateral velocity only affects lateral velocity. The eigenvalue corresponds
to the percent of the perturbation remaining after each stride. DF. Asymptotic recovery in a coupled systemlateral velocity perturbation. In this model, the perturbation in lateral
velocity affects both lateral and rotational velocity. Fast and slow components of recovery are apparent. GI. Asymptotic recovery in a coupled systemdirection of slow recovery.
Perturbation in lateral and rotational velocity. JL. Asymptotic recovery in a coupled systemdirection of fast recovery. Perturbation in lateral and rotational velocity. MO. Neutral
stability in a coupled system. Perturbation in lateral velocity with recovery along the fast direction to a new equilibrium. PR. Instability in a coupled system.
 QUANTIFYING DYNAMIC STABILITY 153

and direction of travel) and rotational velocity (v; in

yaw). Surprisingly, the template passively self-stabiliz-
es when perturbed (e.g., pushed to the side) during
running without the equivalent of neural reflexes (Fig.
4B). A formal dynamical analysis shows that heading
and forward velocity are neutrally stable (eigenvalues
5 1), whereas body orientation and rotational velocity
are asymptotically stable (eigenvalues , 1).
Exert more control in directions that are neutrally
stable or recover slowly via passive dynamics. The
Lateral Leg Spring Template adopts a stable heading
and velocity after a perturbation, but these differ from
those observed before the perturbation (Schmitt et al.,
2002). Stability analysis of the template predicts that
if the animal is to maintain its compass direction
(heading) and forward velocity after a perturbation,
then it must exert neural feedback to do so. In general,
variables that are near neutral stability or are unstable
will likely require more control by active feedback
with the equivalent of neural reflexes. By more con-
trol, we mean a greater amount of computation, and a
greater degree of timely, sensory information stimu-
lating muscles that demand different phase relation-
ships to produce stability.
Exert less control in directions that rapidly recover
via passive dynamics. By contrast, the naive use of
feedback to control a single state variable that recovers
rapidly by way of natural dynamics can actually result
in a less stable system with a slower recovery to a
perturbation. For example, if rotational velocity and
FIG. 4. Model of a passive, dynamic self-stabilizing system. A.
Lateral Leg Spring model (Schmitt and Holmes, 2000a, b). Three
body orientation are perturbed in one direction, then
legs of an insects stance support tripod are collapsed to a single the natural dynamics will first increase the perturbation
virtual leg spring. The template bounces from side to side as it by overshooting the equilibrium and then spiral to-
moves forward. B. Response to perturbation. The diagram shows wards it in a few strides (Fig. 5A; e.g., start lower left).
the model recovering from a perturbation to a stable gait (y 5 Likewise, if rotational velocity and body orientation
0.2175, d 5 0.14, u 5 0, v 5 20.9295 with parameters k 5 2.25,
m 5 0.0025, I 5 2.04 3 1027, d 5 20.0025, b 5 1), applied at the are perturbed in the other direction, then the natural
beginning of the third step. The graphs show the state variables at dynamics will first increase the perturbation by over-
the beginning of each step. The model is neutrally stable in u and y shooting the equilibrium and then spiral towards it in
(corresponding eigenvalues 5 1), but asymptotically stable in v and a few strides (Fig. 5A; e.g., start upper right). Consider
d; the corresponding eigenvalues are less than one in magnitude.
Note that v changes sign with each step and so is not expected to
a perturbation that increases rotational velocity to the
decay to a straight line as the other state variables do. (From Schmitt right, but body orientation to the left (Fig. 5B; e.g.,
et al., 2002.) start upper left). One simple control strategy would be
to urgently realign the body position in the next stride
using negative feedback. Unfortunately, the natural dy-
recently proposed Lateral Leg Spring Model for ter- namics would begin to recover body rotational veloc-
restrial locomotion in the horizontal plane (Fig. 4A; ity, but would misalign the body orientation nearly
Schmitt and Holmes, 2000a, b), inspired by the more back to the original perturbed value. If negative feed-
anchored model (Kubow and Full, 1999). back were applied repeatedly (Fig. 5B; dashed lines),
Self-stabilization without neural reflexesthe lat- the next stride would consistently result in a re-per-
eral leg spring template. Animals appear to bounce turbation of body orientation. A more effective control
from side to side as they run, just as they bounce up strategy would be to simply do nothing and let the
and down. A single virtual passive leg spring project- natural dynamics correct the perturbation over several
ing laterally in the horizontal plane represents a group strides. Alternatively, negative feedback could be ap-
of legs in each stance phase (Schmitt and Holmes, plied along the eigenvector adjusting rotational veloc-
2000a). The dynamic model has three degrees of free- ity in concert with body orientation.
dom, as the body is free to rotate and translate in the
fore-aft and lateral direction. Relevant state variables Forming hypotheses on the dynamic stability of
describing its behavior include heading (u 1 d; the animal design
compass direction), forward velocity (y), body orien- A formal stability analysis used in conjunction with
tation (d; difference between the long axis of the body a dynamic model can yield testable hypotheses of an-
154 R. J. FULL ET AL.

stability in particular. When parameter values approx-

imating those of an insect (Blaberus discoidalis) are
input into the model, the model self-stabilizes after a
perturbation without need for negative feedback. The
extent of this stability (eigenvalue or percentage of the
perturbation remaining after each cycle) varies as a
function of the parameter values.
Stability varies as forward velocity is increased
(Fig. 6A). At subcritical forward velocities, here less
than 0.16 m/sec, the template predicts that no stable
gaits exist. At forward velocities greater than 0.16 m/
sec, the template predicts a periodic gait with heading
and forward velocity after a perturbation being neu-
trally stable (eigenvalue 5 1), whereas body orienta-
tion and rotational velocity are asymptotically stable
(eigenvalue , 1). The effectiveness of recovery of the
template can be shown best by quantifying the recov-
ery rate of the slowest variable (maximum eigenvalue
less than 1) at each forward velocity. As a parameter
is varied, the template is most stable when its slowest
recovering variable recovers more rapidly than any
other parameter value. With respect to forward veloc-
ity, the template predicts the insect to be most stable
at a velocity of 0.22 m/sec (Schmitt et al., 2002).
Slower or faster velocities are less stable. Remarkably,
the preferred speed of the animal is very near 0.22 m/
sec (maximum speed approximately 0.7 m/sec; Kram
et al., 1997; Ting et al., 1994; Full and Tu, 1990). A
similar pattern is observed for the other parameters.
The most stable virtual leg stiffness value is found near
the estimate for the actual insect (Fig. 6B). Relatively
small variations in inertia away from the animals val-
ue decreases stability (Fig. 6C). Varying virtual leg
length and landing angle (sprawl) away from the ani-
mals values also reduce stability (Fig. 6D, E). Hence,
FIG. 5. Control strategies working with and against the natural dy- the template predicts that the insect is operating at or
namics. A. Return map of the Lateral Leg Spring Template to per- near optimal conditions for passive, dynamic self-sta-
turbations in rotational velocity and body orientation. The star rep-
resents the equilibrium values. Arrows show the path of recovery bilization.
dictated by the natural dynamics with no feedback imposed. Thin Testing hypotheses of passive, dynamic self-stabi-
arrows represent one stride and thick arrows three or more strides. lization in running animals is challenging. Preliminary
B. Negative feedback working against the natural dynamic. The experiments on cockroaches lend support to this hy-
closed circle represents a perturbation away from equilibrium (star)
in the direction of the pointing hand. Dashed arrows show an in-
pothesis for brief (10 msec), lateral perturbations (Jin-
stantaneous correction in body orientation to the equilibrium value drich and Full, 2001). Cockroaches were able to ini-
provided by negative feedback. Thin arrows represent changes in tiate recovery in less than 15 msec within the step
one stride due to the natural dynamics. Recovery takes fewer strides during which an impulsive perturbation occurred. Pre-
without feedback when the natural dynamics is rapid. flexes (musculo-skeletal, spring and damper-like be-
havior) appear to dominate rapid stabilization, rather
than neural reflexes.
imal design when the models parameters are varied.
Small variations in parameters may reflect individual Quantifying maneuverability with simple measures
differences, whereas larger changes could make pre- Maneuverable animals can veer, climb, dive, turn,
dictions about species differences. Varying individual stop and negotiate obstacles with ease. Descriptions of
parameters is likely to reveal just how tuned or de- maneuverability for all modes of locomotion are grow-
pendent parameters are on one another. ing in number (see this volume). Unfortunately, ma-
The Lateral Leg Spring Template predicts that in- neuverability can be far more difficult to quantify fully
sects may be a rather highly tuned system. Average than stability. A common theory and general maneu-
forward velocity (y), body mass (m), virtual leg spring verability metrics need to be developed. Here we de-
stiffness (k), body inertia (I), landing angle (b), and scribe two indices for terrestrial turning as an example
virtual leg length (l) are parameters of the model that of simple measures that might allow valuable compar-
can profoundly affect its behavior in general and its isons (Jindrich and Full, 1999).
 QUANTIFYING DYNAMIC STABILITY 155

FIG. 6. Stability of the Lateral Leg Spring Template as a function of speed, body mass and animal design. The stability measure is the amount
of perturbation remaining after each stride (eigenvalue). Eigenvalues less than one reveal asymptotic stability, whereas as values greater than
one represent instability. Eigenvalues equal to one signify neutral stability. Forward velocity (y) and heading (u) are neutrally stable. Asymp-
totically stable eigenvalues plotted correspond to the slowest recovering variable (maximum among eigenvalues less than 1) at each parameter
value. Arrows represent the parameter value estimate for an insect (cockroach, Blaberus discoidalis). A. Forward velocity (y). Grey area
represents the range of velocities where no stable, periodic gaits are found. B. Virtual leg spring stiffness (k). C. Body inertia (I). D. Virtual
leg length (l). E. Landing angle (b).

Simple maneuverability measures for turning. Exe- LMN value. Running cockroaches are able to generate
cuting a successful turn involves at least two require- perpendicular force impulses comparable to 3/4 their
ments. The animals average heading (the direction of linear momentum (LMN 5 0.75) to produce 2273
the average velocity vector of the center of mass) must degree turns. The LMN has also been used by Drucker
be deflected, and the animals body must rotate to keep and Lauder (2001) to quantify turning in sunfish where
the body axis aligned with its heading (Fig. 7). In me- values ranged from 0.120.74.
chanical terms, the change in linear momentum must The second requirement for a successful turn in-
be accompanied by a transient pulse in angular mo- volves rotating the body axis so that it aligns with the
mentum. Animals must generate coordinated linear direction of the velocity vector (Fig. 7B). Minimizing
and angular impulses. the degree of misalignment between the angle of de-
The amount of deflection or change in heading will flection (ud) and the angle of body rotation (ur) is im-
depend on the impulse perpendicular to the direction portant for resuming straight running following a turn
of movement relative to the animals forward momen- (Jindrich and Full, 1999). The ratio of ur/ud can provide
tum (Linear Maneuverability Number, LMN; Jindrich an index of how effective a propulsor or an appendage,
and Full, 1999). A larger perpendicular force exerted such as a leg, can be in generating a torque to match
over a longer period of time will increase the change the angle of deflection to the angle of rotation (Pro-
in heading or turning angle (high LMN; Fig. 7A). pulsor effectiveness number, PEN). The torque gen-
Greater body mass or faster forward velocity for the erated by a perpendicular force can under-rotate the
same perpendicular impulse will decrease the change body (ur , ud; PEN , 1) or over-rotate it (ur . ud;
in heading and cause a smaller turn indicated by a low PEN . 1). Analysis of cockroach turning shows that
156 R. J. FULL ET AL.

FIG. 7. Maneuverability measures. A. Change in heading. Perpendicular force generating a deflection of angle ud. B. Change in body
orientation. Torque generating a body rotation of angle ur. C. Single leg forces required to produce a turn as a function of leg position. The
grid shows position relative to the center of mass. Arrows represent force vectors necessary to turn model so that the angle of deflection equals
the angle of rotation. (After Jindrich and Full, 1999.)

legs on the outside of the turn placed well in front of is also critical if animal design is to be determined.
the center of mass are most effective in producing Animals appear to be grossly over-built for steady-
turns, whereas hind legs are the least effective and state, straight-ahead locomotion. They appear to pos-
require the largest forces (Fig. 7C). sess too many neurons, muscles, joints and even too
Turning the Lateral Leg Spring Templatealtering many appendages. Bernstein (1935) realized that for
the center of pressure is most effective. Turning in any particular task neuromotor synergies must exist.
many-legged, sprawled-posture animals is most often Groups of neurons, muscles, joints and legs work to-
viewed as resulting from an increase in step length, gether as if they were one (see, for example, Full,
leg length or stride frequency of the outside legs 1993; Raasch et al., 1997). We argue that to ultimately
(Franklin et al., 1981; Frantsevich and Mokrushov, reveal function, an animal performing steady-state lo-
1980; Graham, 1972; Jander, 1985; Strauss and Hei- comotion must be subjected to perturbations (Full and
senberg, 1990; Zollikofer, 1994). Small turns by the Koditschek, 1999). Neuromotor synergies must be dis-
Lateral Leg Spring Template can be elicited in 26 covered, probed and broken if we desire a more com-
steps by lengthening and/or stiffening the outside vir- plete understanding of an organism and all its parts.
tual leg spring or increasing its step length (i.e., de-
creasing its touchdown angle, b; Schmitt and Holmes, ACKNOWLEDGMENTS
2000b). But such turns are fragile. Parameters must be Thanks to Devin Jindrich for reading the manu-
selected carefully to avoid unstable behavior and/or script. Supported by DARPA/ONR N00014-98-1-
unnatural body orientations and headings. Propulsor 0747, ONR N00014-98-1-0669 and DoE DE-FG02-
effectiveness is poor, as the body orientation does not 95ER 25238.
match the changes in heading. Although these mech-
anisms can produce a turn in actual animals, the small REFERENCES
changes in the kinematics of cockroach turning are Alexander, R. M. 1982. Locomotion of Animals. In R. M. Alexander
inconsistent with observed behavior (Jindrich and Full, (ed.), Tertiary level biology, pp. 163. Blackie, Glasow.
1999). Bauby, C. E. and A. D. Kuo. 2000. Active control of lateral balance
in human walking. J. Biomech. 33:1,4331,440.
Large turns of 2070 degrees are more easily elic- Beck, P. J., C. E. Peper, and D. F. Stegeman. 1995. Dyanmic models
ited in the Lateral Leg Spring Template by a forward of movement coordination. Human Movement Sci. 14:573608.
shift of the center of pressure (attachment point of the Bernstein, N. 1935, 1967. The co-ordination and regulation of
outside leg). The same appears to be true in the animal movements. Pergamon, New York.
(Jindrich and Full, 1999). The individual foot forces Blickhan, R. and R. J. Full. 1987. Locomotion energetics of the
ghost crab II. Mechanics of the centre of mass during walking
of cockroaches sum to act as a single leg generating a and running. J. Exp. Biol. 130:155174.
force at a particular point on the body equivalent to Blickhan, R. and R. J. Full. 1993. Similarity in multilegged loco-
the center of pressure. As in the Lateral Leg Spring motion: Bouncing like a monopode. J. Comp. Physiol. A 173:
Template, cockroaches appear to alter their center of 509517.
Cavagna, G. A., N. C. Heglund, and C. R. Taylor. 1977. Mechanical
pressure to produce a turn by using legs on the outside work in terrestrial locomotion: Two basic mechanisms for min-
of the turn to generate the majority of force and torque imizing energy expenditure. Am. J. Physiol. 233:R243R261.
impulse. Thus transient center of pressure positions Cruse, H. 1976. The function of legs in the free walking stick insect,
may provide a third useful turning index. Carausius morosus. J. Comp. Physiol. 112:135162.
Cruse, H. 1990. What mechanisms coordinate leg movement in
The next steps in animal locomotionstability and walking arthropods? TINS 13:1521.
maneuverability Delcomyn, F. 1985. Walking and running. In G. A. Kerkut and L. I.
Gilbert (eds.), Comprehensive insect physiology, biochemistry,
The study of stability and maneuverability not only and pharmacology. Nervous system: Structure and motor func-
has obvious behavioral and ecological relevance, but tion, Vol. 5, pp. 439466. Pergamon Press, New York.
 QUANTIFYING DYNAMIC STABILITY 157

Dickinson, M. H., C. T. Farley, R. J. Full, M. A. R. Koehl, R. Kram, Kram, R., B. Wong, and R. J. Full. 1997. Three-dimensional kine-
and S. Lehman. 2000. How animals move: An integrative view. matics and limb kinetic energy of running cockroaches. J. Exp.
Science 288:100106. Biol. 200:1,9191,929.
Dingwell, J. B. and J. P. Cusumano. 2000. Nonlinear time series Kubow, T. M. and R. J. Full. 1999. The role of the mechanical
analysis of normal and pathological human walking. Chaos 10: system in control: A hypothesis of self-stabilization in hexa-
848863. pedal runners. Phil. Trans. Roy. Soc. London B 354:849862.
Drucker, E. G. and G. V. Lauder. 2001. Wake dynamics and fluid Mackinnon, C. D. and D. A. Winter. 1993. Control of whole body
forces of turning maneuvers in sunfish. J. Exp. Bio. 204:431 balance in the frontal plane during human walking. J. Biomech.
442. 26:633644.
Duysens, J., F. Clarac, and H. Cruse. 2000. Load-regulating mech- McGhee, R. B. 1976. Robot locomotion. In M. Herman, S. Grillner,
anisms in gait and posture: Comparative aspects. Physiological P. S. G. Stein, and D. G. Styart (eds.), Neural control of loco-
Reviews 80(1):83133. motion, pp. 237264. Plenum, New York.
Franklin, R., W. J. Bell, and R. Jander. 1981. Rotational locomotion McIntyre, J. and E. Bizzi. 1993. Servo hypotheses for the biological
by the cockroach Blattella germanica. J. Insect Physio. 27:249 control of movement. J. Motor Behav. 25:193202.
255. Pearson, K. G. 1993. Common principles of motor control in ver-
Frantsevich, L. I. and P. A. Mokrushov. 1980. Turning and righting tebrates and invertebrates. Annu. Rev. Neurosci. 16:26597.
in Geotrupes (Coleoptera, Scarabeidae). J. Comp. Physiol. 136: Raasch, C. C., F. E. Zajac, B. M. Ma, and W. S. Levine. 1997.
279289. Muscle coordination of maximum-speed pedaling. J. Biome-
Full, R. J. 1993. Integration of individual leg dynamics with whole chanics 30:595602.
body movement in arthropod locomotion. In R. Beer, R. Ritz- Schmitt, J. and P. Holmes. 2000a. Mechanical models for insect
mann, and T. McKenna (eds.), Biological neural networks in locomotion: Dynamics and stability in the horizontal plane I.
invertebrate neuroethology and robots, pp. 320. Academic Theory. Biol. Cybern. 83:501515.
Press, Boston. Schmitt, J. and P. Holmes. 2000b. Mechanical models for insect
Full, R. J. and D. E. Koditschek. 1999. Templates and anchors: Neu- locomotion: Dynamics and stability in the horizontal plane II.
romechanical hypotheses of legged locomotion on land. J. Exp. Application. Biol. Cybern. 83:517527.
Biol. 202:3,3253,332. Schmitt, J., M. Garcia, R. C. Razo, P. Holmes, and R. J. Full. 2002.
Full, R. J. and M. S. Tu. 1990. Mechanics of six-legged runners. J. Dynamics and stability of legged locomotion in the horizontal
Exp. Biol. 148:129146. plane: A test case using insects. Bio. Cyb. (In press).
Full, R. J. and M. S. Tu. 1991. Mechanics of a rapid running insect: Strogatz, S. H. 1994. Nonlinear dynamics and Chaos. Addison Wes-
Two-, four- and six-legged locomotion. J. Exp. Biol. 156:215 ley, Reading, Massachusetts.
231. Song, S. M. 1984. Kinematic optimal design of a six-legged walking
Graham, D. 1972. A behavioural analysis of the temporal organi- machine. Ph.D. Diss. Ohio State University, Columbus, Ohio.
sation of walking movements in the 1st instar and adult stick Sun, S. S. 1974. A theoretical study of gaits for legged locomotion
systems. Ph.D. Diss. Ohio State University, Columbus, Ohio.
insect (Carausius morosus). J. Comp. Physiol. 81:2352.
Strauss, R. and M. Heisenberg. 1990. Coordination of legs during
Graham, D. 1985. Pattern and control of walking in insects. Adv.
straight walking and turning in Drosophila melanogaster. J.
Insect Physiol. 18:31140.
Comp. Physiol. A 167:403412.
Guckenheimer, J. and P. Holmes. 1983. Non-linear oscillations, dy-
Taga, G. 1995a. A model of the neuro-musculo-skeletal system for
namical systems, and bifurcations of vector fields. Springer-Ver- human locomotion 1. Emergence of basic gait. Bio. Cyber. 73:
lag, New York. 97111.
Hanken, H., J. A. S. Kelos, and H. Bunz. 1985. A theoretical model Taga, G. 1995b. A model of the neuro-musculo-skeletal system for
of phase transitions in human hand movements. Bio. Cyber. 51: human locomotion II. Real-time adaptability under various con-
347356. straints. Bio. Cyber. 73:113121.
Hogan, N. 1990. Mechanical impedance of single- and multi-artic- Taga, G., Y. Yamaguchi, and H. Shimizu. 1991. Self-organized con-
ular systems. In J. A. W. Winters, S. L. (ed.), Multiple muscle trol of bipedal locomotion by neural oscillators in unpredictable
systems: Biomechanics and movement organization, pp. 149 environments. Bio. Cyber. 65:147159.
164. Springer-Verlag, New York. Taylor, C. R., N. C. Heglund, and G. M. O. Mailoy. 1982. Energetics
Hughes, G. M. 1952. The co-ordination of insect movements. I. The and mechanics of terrestrial locomotion. I. Metabolic energy
walking movements of insects. J. Exp. Biol. 29(2):267283. consumption as a function of speed and body size in birds and
Hughes, G. M. 1957. The co-ordination of insect movements. II. the mammals. J. Exp. Bio. 97:121.
effect of limb amputation and the cutting of commissures in the Taylor, C. R. and E. R. Weibel. 1981. Design of the mammalian
cockroach (Blatta oreientalis). J. Exp. Biol. 34:306333. respiratory system. I. Problem and Strategy. Resp. Physiol. 44:
Jander, J. P. 1985. Mechanical stability in stick insects when walking 110.
straight and around curves. In M. Gewecke and G. Wendler Ting, L. H., R. Blickhan, and R. J. Full. 1994. Dynamic and static
(eds.), Insect locomotion, pp. 3342. Paul Parey, Berlin. stability in hexapedal runners. J. Exp. Biol. 197:251269.
Jindrich, D. L. and R. J. Full. 1999. Many-legged maneuverability: Zehr, E. P. and R. B. Stein. 1999. What functions do reflexes serve
dynamics of turning in hexapods. J. Exp. Bio. 202:1,6031,623. during human locomotion? Prog. Neurobiol. 58:185205.
Jindrich, D. L. and R. J. Full. 2001. Dynamic stabilization of rapid Zollikofer, C. P. 1994. Stepping patterns in ants I. Influence of speed
hexapodal locomotion. Amer. Zool. 40:247. and curvature. J. Exp. Biol. 192:95106.