ME 328: Medical Robotics

Autumn 2016

Lecture 5:
Robot dynamics and simulation

Allison Okamura
Stanford University
Robot dynamics
 equations of motion
describe the relationship between forces/torques and
motion (in joint space or workspace variables)

two possible goals:

~ ~˙ ¨
1. Given motion variables (e.g. ✓, ✓, ✓ or ~x, ~x˙ , ~x¨ ), what
~
joint torques ( ~⌧ ) or end-effector forces ( f~ ) would
have been the cause? (this is inverse dynamics)
~
2. Given joint torques ( ~⌧ ) or end-effector forces ( ), f
~ ~˙ ¨
what motions (e.g. ✓, ✓, ✓ or ~x, ~x˙ , ~x¨ ) would result?
~
(this is forward dynamics)
developing equations of motion
using Lagrange’s equation
The Lagrangian is L=T V
where T is the kinetic energy of the system and V is
the potential energy of the system
✓ ◆
d @L @L
Lagrange’s equation is = Qj
dt @ q̇j @qj

@qj
where j = 1, 2, . . . , n , and q̇j = is the generalized
@t
velocity and j is the nonconservative generalized
Q
force corresponding to the generalized coordinate qj
what are generalized coordinates?
• equations of motion can be formalized in a
number of different coordinate systems

• n independent coordinates are necessary to

describe the motion of a system having n degrees
of freedom

• any set of n independent coordinates is called

generalized coordinates: q1 , q2 , . . . qn

• these coordinates may be lengths, angles, etc.

 generalized forces
• When external forces act on the system, the
configuration changes: generalized coordinates qj
change by qj , j = 1, 2, . . . , n

• If Uj is the work done in changing qj by qj , the

Uj
corresponding generalized force is Qj = ,
q
where j = 1, 2, . . . , n j

• Qj is a force/moment and qj is a linear/angular

displacement. This can include dissipation (damping).
d @L @L
dt ( @ q̇j ) = Qj , where j = 1, 2, . . . , n
where does Lagrange’s
@qj

@qj
where q̇ = j is the generalized velocity and Q is the nonconservative
j

equation come from?

@t
generalized force corresponding to the generalized coordinate q . j

Where does it come from?

Hamilton’s principle of least action: a system moves from q(t1 ) to q(t2 )

in such
R t2 a way that the following integral takes on the least possible value.
S = t1 L(q, q̇, t)dt

The calculus of variations is used to obtain Lagrange’s equations of mo-

tion.
R t2 We’re concerned with minimizing
t1 f (y(t), ẏ(t); t) dt

The minimization leads to the equation

@f d @f
@y dt @ ẏ = 0

If there is more than one set of variables in the functional f (e.g. yi and ẏi )
then you get one equation for each set.
T = 2 q̇ M q̇

using Lagrange’s equation to

M ! generalized mass matrix
Note: In general, M need not be diagonal

derive equations of motion

Using Lagrange’s Equation to derive Equations of Motion
d @L @L
dt ( @ q̇j ) @qj = Qj , where j = 1, 2, . . . , n

L=T V

Substitute:
@L @T @V
@ q̇j = @ q̇j @ q̇j
@L @T
Last term is zero because P.E. is not dependent on velocities ! @ q̇j = @ q̇j
@L @T @V
@qj = @qj @qj

d @T @T @V
dt ( @ q̇j ) @qj + @qj = Qj , where j = 1, 2, . . . , n

Qj is a nonconservative generalized force corresponding to coordinate qj ,

e.g. damping
For a conservative system,
d @T @T @V
(
dt @ q̇j ) @qj + @qj = 0, where j = 1, 2, . . . , n
 adding dissipation
since the left side of Lagrange’s equation only includes
terms for potential and kinetic energy, any dissipative
terms (e.g., damping) must be added on the right hand
side (and Qj are now only the input forces/torques)

is a simplified form of Rayleigh’s dissipation function

example: double pendulum
(review on your own)
Velocity of m1 : v1 = l1 ✓1
1
2
Velocity of m2 : v2 = (v2x + v2y ) 2 2

example: double pendulum

v2x = l1 ✓˙1 cos(✓1 ) + l2 ✓˙2 cos(✓2 )
v2y = l1 ✓˙1 sin(✓1 ) + l2 ✓˙2 sin(✓2 )
Velocity of m1 : v1 = l1 ✓˙1
1
2 2
Kinetic energy: Velocity of m2 : v2 = (v2x + v2y ) 2
˙1 cos(✓ ˙2 2cos(✓2 ) ˙
T = 2 m1 (l1 ✓1 ) + 2 m2 ((l1 ✓1 cos(✓1 )+l˙ 2 ✓2 cos(✓2 ˙)) +(l1 ✓1 sin(✓1 )+l2 ✓˙2 sin(✓2 ))2 )
1 ˙ 2 1 ˙ v 2x = l 1 ✓ ˙ 1 ) + l 2 ✓
v2y = l1 ✓1 sin(✓1 ) + l2 ✓2 sin(✓2 )

Kinetic energy:
Potential energy
T = 12 m1 (l1 ✓˙1 )2 + 12 m2 ((l1 ✓˙1 cos(✓1 )+l2 ✓˙2 cos(✓2 ))2 +(l1 ✓˙1 sin(✓1 )+l2 ✓˙2
V = m1 gl1 (1 cos(✓1 )) + m2 g(l1 (1 cos(✓1 )) + l2 (1 cos(✓2 )))

The Lagrangian is:

L=T V
L = 1
(m +m )l 2 ✓˙ 2 + 1 m l2 ✓˙ 2 +m
Potential
l l ˙
✓ ˙2 cos(✓1 ✓2 )+(m1 +m2 )gl1 cos(✓1 )+
energy
✓
Potential
2 1 energy
2 1 1 2 2 2 2 2 1 2 1
V = m1 gl1 (1 cos(✓1 )) + m2 g(l1 (1 cos(✓1 )) + l2 (1 cos(✓2 )))
Vm2=glm
2 cos(✓ 2)
1 gl1 (1 cos(✓1 )) + m2 g(l1 (1 cos(✓1 )) + l2 (1 cos(✓2 )))
The Lagrangian is:

example: double pendulum

L=T V
L = 12 (m1 +m2 )l12 ✓˙12 + 12 m2 l22 ✓˙22 +m2 l1 l2 ✓˙1 ✓˙2 cos(✓1 ✓2 )+(m1 +m2 )gl1 cos(✓1 )+
m2 gl2 cos(✓2 )

For ✓1 :
@L
= m l 2 ✓˙ + m l2 ✓˙ + m l l ✓˙ cos(✓ ✓2 )
@ ✓˙1 1 1 1 2 1 1 2 1 2 2 1
⇣ ⌘
d @L
= (m1 +m 2 )l 2✓¨1 +m2 l1 l2 ✓¨2 cos(✓1 ✓2 ) m2 l1 l2 ✓˙2 sin(✓1 ✓2 )(✓˙1 ✓˙2 )
dt @ ✓˙ 1
1

@L
@✓1 = l1 g(m1 + m2 ) sin(✓1 ) m2 l1 l2 ✓˙1 ✓˙2 sin(✓1 ✓2 )

Thus, the di↵erential equation for ✓1 becomes:

(m1 + m2 )l12 ✓¨1 + m2 l1 l2 ✓¨2 cos(✓1 ✓2 ) + m2 l1 l2 ✓˙22 sin(✓1 ✓2 ) + l1 g(m1 +
m2 ) sin(✓1 ) = 0

Divide through by l1 and this simplifies to:

(m1 +m2 )l1 ✓¨1 +m2 l2 ✓¨2 cos(✓1 ✓2 )+m2 l2 ✓˙22 sin(✓1 ✓2 )+g(m1 +m2 ) sin(✓1 ) =
0

Similarly, for ✓2 :
@L
= m l 2 ✓˙ + m l l ✓˙ cos(✓ ✓2 )
@ ✓˙ 2 2 2 2 1 2 1 1
Divide through by l1 and this simplifies to:
example: double pendulum
(m1 +m2 )l1 ✓¨1 +m2 l2 ✓¨2 cos(✓1 ✓2 )+m2 l2 ✓˙22 sin(✓1 ✓2 )+g(m1 +m2 ) sin(✓1 ) =
0

Similarly, for ✓2 :
@L
= m l 2 ✓˙ + m l l ✓˙ cos(✓ ✓2 )
@ ✓˙2 2 2 2 2 1 2 1 1
⇣ ⌘
d @L
= m 2 l 2
¨2 + m2 l1 l2 ✓¨1 cos(✓1 ✓2 )
✓ m2 l1 l2 ✓˙1 sin(✓1 ✓2 )(✓˙1 ✓˙2 )
dt @ ✓˙ 2

@L
@✓2 = m2 l1 l2 ✓˙1 ✓˙2 sin(✓1 ✓2 ) l2 m2 g sin ✓2

Thus, the di↵erential equation for ✓2 becomes:

m2 l22 ✓¨2 + m2 l1 l2 ✓¨1 cos(✓1 ✓2 ) m2 l1 l2 ✓˙12 sin(✓1 ✓2 ) + l2 m2 g sin ✓2 = 0

Divide through by l2 and this simplifies to:

m2 l2 ✓¨2 + m2 l1 ✓¨1 cos(✓1 ✓2 ) m2 l1 ✓˙12 sin(✓1 ✓2 ) + m2 g sin ✓2 = 0

So, we have developed a very complicated set of coupled equations of motion

from a very simple system!

⇡
For the initial conditions ✓1 = 2 and ✓2 = ⇡, you get the response be-
but robots are not point masses...
links could be represented by solid cylinders

moments of inertia about center:

2
mr
Iz =
2
1 2 2
Ix = Iy = m 3r + h
12
moment of inertia tensor:

http://en.wikipedia.org/wiki/List_of_moments_of_inertia
what do you do if the rotation is not
about the center?
If the new axis of rotation is parallel to the original
axis of rotation, use the parallel axis theorem:
2
Inew = Icm + mR
Icm moment of inertia of the object about an
axis passing through its center of mass
m object’s mass
R perpendicular distance between the two axes
 and what if the new axis of rotation is
not parallel to the original?
the moment of inertia of a continuous solid body
rotating about a known
Z axis is calculated by
I= ⇢(~r)d(~r)dV (~r)
V
~r radius vector (from origin to volume element of interest)
⇢(~r) object’s mass density at ~r
d(~r) shortest distance between a point at ~r and the axis
of rotation
integrated over the volume V of the body

... but to reduce the complexity of this week’s assignment, we will

approximate the moment of inertia by assuming that we do have parallel axes
considering kinetic energy of solids
sum translational and rotational components
1 2 1 2
T = mv + I!
2 2

http://hyperphysics.phy-astr.gsu.edu/hbase/rke.html
 for robot dynamics review
http://see.stanford.edu/materials/aiircs223a/
handout7_Dynamics.pdf

http://cs.stanford.edu/groups/manips/teaching/cs223a/
assets/slides/dynamics.pdf

... and many dynamics and robotics textbooks

In Assignment 3, we will give you the dynamic equations,

but it helps to understand where they come from!
 questions
• how do you think an RCM robot compares
to a typical serial chain manipulator in
terms of its dynamics?

• does the da Vinci have haptic feedback?

(and how do the system dynamics affect
this capability?)
Dynamic simulation
 controller on one end,
system dynamics on the other
a controller computes
the desired force
e.g. f = kp*(x-xd) desired force
(in computer)

endpoint
force/torque
in Assignment 3,
you will simulate the this force and externally applied
effects of system loads result in robot motion
dynamics e.g., solve for x in f = mẍ + bẋ
 simulating equations of motion
goal: given an equation of motion and applied forces,
what will the resulting robot motion be?

The dynamics equations we have are coupled, non-linear

ODEs. They are hard (likely impossible) to solve analytically.

Instead, we solve them using numerical integration.

you can do a simple calculation yourself,

i.e. integration using the trapezoidal rule, or use a handy
and more accurate/robust Matlab function
(Simulink is also an option)
plotting and data manipulation. However, Matlab can also be a powerful tool for solving ordinary
di↵erential equations (ODEs) directly.

There are a series of native Matlab functions that solve ODEs given initial conditions. Di↵erent
functions are appropriate for di↵erent problem types (“sti↵ness”, corresponding to large di↵erences
in scales) and and order of accuracy (low to high).

Solver Problem Type Order of Accuracy When to Use

ode45 Nonsti↵ Medium Most of the time. This should be the first
solver you try.
ode23 Nonsti↵ Low For problems with crude error tolerances or
for solving moderately sti↵ problems.
ode113 Nonsti↵ Low to high For problems with stringent error tolerances
or for solving computationally intensive prob-
lems.
ode15s Sti↵ Low to medium If ode45 is slow because the problem is sti↵.
ode23s Sti↵ Low If using crude error tolerances to solve sti↵
systems and the mass matrix is constant.
ode23t Moderately Sti↵ Low For moderately sti↵ problems if you need a
solution without numerical damping.
ode23tb Sti↵ Low If using crude error tolerances to solve sti↵
systems.

ode45 is based on an explicit Runge-Kutta (4,5) formula, the Dormand-Prince pair. It is a one-step
solver - in computing y(tn ), it needs only the solution at the immediately preceding time point,
y(tn-1). In general, ode45 is the best function to apply as a “first try” for most problems.

The basic syntax for these solvers is:

 where:
The basic syntax for these solvers is:
e45,[T,Y] isetc.
one of ode45,
= solver(odefun,tspan,y0)
solver
ode23, ode23, etc.
handle
odefun
where: thatis evaluates
a functionthe handle
rightthat side evaluates the rightequat
of the di↵erential side
solver
ecifying
tspan isthe
one
is of
a ode45,
vector
interval ode23, etc.
specifying
of integration, the [t0,tf]
interval of integration, [t
odefun is a function handle that evaluates the right side of the di↵erential equations
y0 isis aavector
ial tspan
conditions vector of initial
specifying conditions
the interval of integration, [t0,tf]
y0 is a vector of initial conditions

Example: First order system

Example:
order system First order system
Let’s say that we want to solve the function
Let’s
ant to saythe
solve that we want to solve
function the
ẏ + 2y = 0
function
for the initial condition y(0) = 10.
ẏ + 2y = 0 ẏ + 2y = 0

ion for the

y(0) = initial
10. condition y(0) = 310.
 First, you need to create a function of the form ẏ = f (t, y):
function ydot = my_function(t,y)

ydot = -2*y; % end of subprogram

First, you need to create a functionAnd
ofthethe form ẏ = f (t, y):
code to apply initial condition and solve the system is:
function ydot = my_function(t,y) tspan = [0,5]; % time duration for calculation
y0 = 10; % initial condition

% Note that the equation of motion is in a

ydot = -2*y; % end of subprogram
% subprogram named my_func

[t,y] = ode45(’my_func’,tspan,y0);
And the code to apply initial condition and solve the system is:
% plot the response
tspan = [0,5]; % figure
time duration for calculation
plot(t,y)
y0 = 10; % initial condition title(’First Order Response’)
xlabel(’Time (s)’)
ylabel(’Position (m)’)
% Note that the equation of motion is in a
The resulting plot is shown below:
% subprogram named my_func First Order Response
10

9
[t,y] = ode45(’my_func’,tspan,y0); 8

% plot the response 6

Position (m)
figure 5

plot(t,y) 4

title(’First Order Response’) 3

2
xlabel(’Time (s)’)
1

ylabel(’Position (m)’) 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s)

The resulting plot is shown below:

 trajectory generation

reference: Chapter 7 of Introduction to Robotics by J. J. Craig (any edition)

 discussion

why would you want to generate a trajectory?

would you want to do trajectory generation in

Cartesian space or joint space?
teleoperation autonomy
surgical planning
user provides the
specifies the start/
trajectory at each
end points and
control loop
desired via points

for autonomous robots, we typically specify the

trajectory based on start point, end point, via points,
motion (e.g., velocity) constraints,
and/or time constraints
 discussion

what properties might you want in a trajectory?

 Trajectory generation goal
move a manipulator from an initial pose to a
final pose in a smooth manner

what does smooth mean?

at least C1 continuous position profile
at least C0 continuous velocity profile
possibly continuous acceleration profile
point-to-point trajectory generation
consider the problem of moving a robot end-
effector from its initial 3D pose to a final 3D pose
in a certain amount of time

2 3
x(t)
p(0) p(tf )
p(t) = 4 y(t) 5
z(t)

robot graphic courtesy Fidel Hernandez

 a smooth path
2 3
x(t)
cubic polynomial: p(t) = 4 y(t) 5 = a0 + a1 t + a2 t2 + a3 t3
z(t)

x(t) = ax0 + ax1 t + ax2 t2 + ax3 t3

or y(t) = ay0 + ay1 t + ay2 t2 + ay3 t3

z(t) = az0 + az1 t + az2 t2 + az3 t3

2 3
ẋ(t)
derivatives: ṗ(t) = 4 ẏ(t) 5 = a1 + 2a2 t + 3a3 t2
ż(t)
2 3
ẍ(t)
p̈(t) = 4 ÿ(t) 5 = 2a2 + 6a3 t
z̈(t)
 constraints
position and velocity at initial and final times:
p(0) = p0
p(tf ) = pf
ṗ(0) = 0
ṗ(tf ) = 0

with four equations and four unknowns, you can

solve for the coefficients a in terms of p0 and pf
and the time tf
you can pick the time tf based on how quickly you
want to move between to the two points, or based
on a maximum velocity constraint
example
Example
trajectory
A single-link robot with a rotary joint
• Must start and stop at rest
position
• Must start from θ0 = 15° and stop at θf = 75°
•cubic
Must do polynomial
so in tf = 3 seconds
C1 continuous
Position
θ ( t ) = 15 .00 + 20 . 00 t 2 + − 4 . 44 t 3

velocity
quadratic
Velocitypolynomial
C0 continuous
.
θ (t ) = 40.00t + 13.33t 2

Acceleration
acceleration
..
linear θpolynomial
( t ) = 40 . 00 − 26 . 66 t

not continuous (at t = 0)

 now what?

now that you have created a trajectory

(i.e., you know the desired position
of the robot at each time step),
you can control the robot to follow
this trajectory

previously, we generated this trajectory

using the “master” of the teleoperator
 discussion

why is a smooth trajectory important?

why might you want a smoother trajectory?
how could you compute this?

what do the produced trajectories look like spatially? what

if you design in joint space instead of Cartesian space?

how might via points be useful?

how would you generate trajectories for them?
 discussion
how would you compute tf based on a
defined maximum velocity?

what kind of trajectory would you want for a

robot that inserts a needle into solid tissue?

bloodbot (Imperial College of London)

 Overview
g, similar to dynamic model or
user input or Assignment 2
trajectory physical robot

xd(t) forces x(t)

PD
+- plant
controller

x’(t) = f(x(t), gpd(x(t), xd(t)))

the ODE we solve numerically

 Assignment 3
Problem 1: Commentary on seminar
Problem 2: Read papers, answer questions
Problem 3: Modeling and simulation of medical
robot dynamics
Problem 4: Effects of dynamics on robot control

Due Wednesday October 19