Jacobians in the Force Domain

Craig Carignan
Glen Henshaw
October 8, 2023

Abstract

1 Intro
Let’s consider moment and force propagation through a link, instead of angular
and linear velocity propagation.
In the static case:
y
4

3 f

2 τ

1 τ

0
0 1 2 3 4 5 6 x

What τ is required to counteract f ? Since this is a statics analysis, we’ll

treat the robot as a structure, locking all of the joints.
Some definitions:

fi force exerted on link i by link i-1

ni torque exerted in link i by link i-1

Since this is a static analysis, we assume that the forces and the torques
must add to zero. Specifically,
i
fi − i fi+1 = 0

1
 ni+1
{i + 1}

{i}
i
Pi+1

ni
fi+1
fi

and
i
ni − i ni+1 − i Pi+1 × i fi+1 = 0
In order to recursively calculate the forces and torques exerted on each link, we
have to start with the last link and work backwards, towards the base frame.
And we need equations that are formulated entirely in terms of the forces and
torques exerted by the distal link. These can easily be seen to be
i i
fi = fi+1 (1)
i i
ni = ni+1 + i Pi+1 × i fi+1 (2)

We’d really like to express the forces and torques entirely in their “own” frames,
so we’ll rotate them as such:
i i i+1
fi = i+1 R fi+1 (3)
i i i+1 i i
ni = i+1 R ni+1 + Pi+1 × fi+1 (4)

NOTE that all forces and moment vectors are absorbed by the structure of
the arm except for those exerted about the axis of rotation of the joint. So to
balance the force/moment equation, we have to isolate the z–component. If the
joint is revolute, this is τi = i nTi i Ẑi . If the joint is prismatic, this is τi = i fiT i Ẑi .

1.1 Virtual Work

The concept here is that work can be expressed in terms of either motion of the
joints or motion of the end effector, and that these two terms must be equal.
Work is the inner product of a vector force or torque with a vector displacement.
In the case of our diagram, this is

F · δX} = τ| ·{zδΘ}
| {z
Cartesian Joint
space space

where F is a Cartesian force–moment vector acting on the end effector, δΘ is

a infinitesimal Cartesian displacement of the end effector, τ is the vector of
torques at the joints, and δΘ is an infinitesimal vector of joint displacements.

2
 You can write a dot product as a vector multiplication, so
T T
F
| {zδX} = τ| {zδΘ}
Cartesian Joint
space space

Recall that the Jacobian relates infinitesimal motions of the joints to infinitesi-
mal motions of the end effector, e.g. δX = JδΘ. So we can write

F T JδΘ = τ T δΘ

And this has to be true for any value of δΘ, so we can cancel this term out to
get
FT J = τT
or
τ = JT F

2 Example

y
6

5
f

3 τ2
θ2

1 τ1
θ1
0
0 1 2 3 4 x

Given f , we want to find τ . From our previous result, we know that

τ = 0J T 0f = 3J T 3f

Normally, we get 3 f from a force–torque sensor attached to the end effector,

which is in frame 3. We get 3 J from Equation 5.66. So:
 T  3 
l1 s2 0 fx
τ= 3
l1 c2 + l2 l1 fy

3
or
τ1 = l1 s2 fx + (l2 + l1 c2 )fy
τ2 = l2 f y

3 Cartesian transformation of velocities and forces

Sometimes we’ll want to consider generalized force/moment and translational/rotational
velocity vectors, e.g.  
v
V=
ω
and  
f
F=
n
We already have all the machinery we need to write 6×6 matrix transformations
to handle these generalized cases. Specifically, we can rewrite Equations 5.45,
5.47 (for revolute joints):
i+1 i+1 i
ωi+1 = i R ωi + θ̇i+1 i+1 Ẑi+1
i+1 i+1
R i vi + i ωi × i Pi+1


vi+1 = i

or equation 5.48 (for prismatic joints):

i+1 i+1 i
ωi+1 = i R ωi
i+1 i+1
R i vi + i ωi × i Pi+1 + d˙i+1 i+1 Ẑi+1


vi+1 = i

IMPORTANT NOTE: we are dealing here with a static situation — we aren’t

considering motion of the robot’s joints here — so we can simplify this as
i+1 i+1 i
ωi+1 = i R ωi
i+1 i+1
R i vi + i ωi × i Pi+1


vi+1 = i

and
i+1 i+1 i
ωi+1 = i R ωi
i+1 i+1
R i vi + i ωi × i Pi+1


vi+1 = i

By factoring these results and integrating them into a single matrix–vector

equation, we get (for revolute joints):
 B   B B A
 A 
vB
= A R − A R PB,Org× vA
B B A
ωB 0 AR ωA
for two frames {A} = i and {B} = i + 1, where we have expressed the cross
product operator in matrix form:
 
0 −pz py
P× =  pz 0 −px 
−py px 0

4
Note also that we have made use of the identity ω × p = −p × ω in the (1, 2)
element of this matrix expression.
We will refer to the matrix above as a velocity transformation Tv :
B B A
VB = A Tv VA .

Be careful not to confuse this with a generalized coordinate transformation

matrix, which is often denoted as i+1
i T !
In exercise 5.1 you are asked to derive the expression for Tv−1 , which is
 A   A
R PB,Org × A RA
 B 
vA B B vB
A = A B
ωA 0 BR ωB
And similarly, from Equations 5.80 and 5.81, we can express the generalized
force/moment transformation as
 A   A
 B 
fA B R 0 fB
A = A
nA PB,Org × A
BR BR
A B
nB
We will refer to the matrix above as a force–moment transformation Tf :
A A B
FA = B Tf FB .

3.1 Example

{T }
ni

fi
{W } Tool

{S}

Sensor

It’s often the case that we will have a force–torque sensor mounted some-
where in the wrist of a robot, but that what we are really interested is the forces
and torques applied at the tool tip, not those in the sensor frame. To transform
from sensed forces and torques to the tool frame, we calculate TS Tf :
T
 
T S R 0
S Tf = T
PS,Org × TS R TS R
and therefore
T T S
FT = S Tf FS