Robotics 1

Differential kinematics
Prof. Alessandro De Luca

Robotics 1

Differential kinematics
!

relations between motion (velocity) in joint space

and motion (linear/angular velocity) in task space
(e.g., Cartesian space)
instantaneous velocity mappings can be obtained
through time derivation of the direct kinematics or
in a geometric way, directly at the differential level
!
!

different treatments arise for rotational quantities

establish the link between angular velocity and
! time derivative of a rotation matrix
! time derivative of the angles in a minimal
representation of orientation

Robotics 1

Angular velocity of a rigid body

vP1
P1

vP3 - vP1

r12

r13
vP3

rigidity constraint on distances among points:

!rij! = constant

vP2 - vP1

P2

vP2
r23

P3
" P1, P2, P3

vPi - vPj orthogonal to rij

1

vP2 - vP1 = !1 ! r12

vP3 - vP1 = !1 ! r13

vP3 - vP2 = !2 ! r23

2-1=3

vPj = vPi + ! ! rij = vPi + S(!) rij

!1 = !2 = !
.
rij = ! ! rij

" the angular velocity ! is associated to the whole body (not to a point)
" if # P1, P2 with vP1=vP2=0: pure rotation (circular motion of all Pj line P1P2)
" !=0: pure translation (all points have the same velocity vP)
Robotics 1

Linear and angular velocity

of the robot end-effector
.
!1 = z01

.
!n = zn-1n

.
!2 = z12
.
v3 = z2 d3

.
!i = zi-1i

alternative definitions
of the direct kinematics
of the end-effector

r = (p, $)
T=

R p
000 1

v and ! are vectors, namely are elements of vector spaces

!
!

they can be obtained as the sum of single contributions (in any order)
these contributions will be those of the single the joint velocities

on the other hand, $ (and d$/dt) is not an element of a vector space

!

a minimal representation of a sequence of two rotations is not obtained summing

the corresponding minimal representations (accordingly, for their time derivatives)

in general, ! % d$/dt
Robotics 1

Finite and infinitesimal translations

!

finitex,y,z or infinitesimal dx, dy, dz translations

(linear displacements) always commute

z
z

y
x

x
z

Robotics 1

same final
position

Finite rotations do not commute

example

z
$ X = 90

initial
orientation

x
z

$ Z = 90
x

mathematical fact: ! is
NOT an exact differential form
(the integral of ! over time
depends on the integration path!)

$ Z = 90

$ X = 90
x

Robotics 1

different final
orientations!

note: finite rotations still commute

when made around the same fixed axis

Infinitesimal rotations commute!

!

infinitesimal rotations d$X, d$Y, d$Z around x, y, z axes

RX($X) =

1
0
0

0
0
cos $X -sin $X
sin $X cos $X

cos $Y 0 sin $Y
RY($Y) =
0
1
0
-sin $Y 0 cos $Y
cos $Z -sin $Z
RZ($Z) = sin $Z cos $Z
0
0
!

&0
&0
1

Robotics 1

RY(d$Y) =

1
0
0 1
-d$ Y 0

RZ(d$Z) =

R(d$) = R(d$X, d$Y, d$Z) =

in any order

RX(d$X) =

1
0
0

1 -d$z d$Y
d$z 1 -d$X
-d$Y d$X 1

= I + S(d$)

0
0
1 -d$X
d$X
1
d$ Y
0
1

1 -d$ Z 0
d$ Z 1 0
0
0
1
neglecting
second- and
third-order
(infinitesimal)
terms
7

Time derivative of a rotation matrix

" let R = R(t) be a rotation matrix, given as a function of time
" since I = R(t)RT(t), taking the time derivative of both sides yields
0 = d[R(t)RT(t)]/dt = dR(t)/dt RT(t) + R(t) dRT(t)/dt
= dR(t)/dt RT(t) + [dR(t)/dt RT(t)]T
thus dR(t)/dt RT(t) = S(t) is a skew-symmetric matrix
" let p(t) = R(t)p a vector (with constant norm) rotated over time
" comparing
dp(t)/dt = dR(t)/dt p = S(t)R(t) p = S(t) p(t)
dp(t)/dt = !(t) ! p(t) = S(!(t)) p(t)
we get S = S(!)
.

R = S(!) R

Robotics 1

!
p

.
p

S(!) = R RT
8

Example

Time derivative of an elementary rotation matrix

RX($(t)) =

1
0
0

0
0
cos $(t) -sin $(t)
sin $(t) cos $(t) &

.
.
T
RX($) R X($) = $ &
0
= 0
0

0
0
0
0 - sin $ - cos $
0 cos $ - sin $ &
0
0.
0. - $
$
0

= S(!)

!=

Robotics 1

1
0
0
0 cos $ sin $
0 - sin $ cos $ &

.
$
0
0&
9

Time derivative of RPY angles and !

RRPY ('x, (y, )z) = RZYX ()z, (y, 'x)
z

the three
contributions
. .
.
)z, (y, 'x to !
are simply
summed as
vectors

x )&

TRPY ((,))

!=
.
)&

.
(&

y
y

(&
.
'&

x
x

c( c) -s)
c( s)& c)&
-s(
0
x
y

1st col in

0
0&
1
z

.
'&
.
(&
.
)&

2nd col in

RZ()z)RY((y) RZ()z)

det TRPY ((,)) = c( = 0

for ( = * /2
(singularity of the
RPY representation)

similar treatment for the other 11 minimal representations...

Robotics 1

10

Robot Jacobian matrices

!

analytical Jacobian (obtained by time differentiation)

r=

p
$

= fr(q)

.
.
+fr(q) .
p
q = Jr(q) q
. =
+q
$

geometric Jacobian (no derivatives)

v
!

Robotics 1

.
r=

.
p
!

JL(q)
JA(q)

.
.
q = J(q) q

11

Analytical Jacobian of planar 2R arm

y
py

l2
l1

direct kinematics
px = l1 c1 + l2 c12

q2
r

q1
px

$ = q1 + q2

.
.
.
.
px = - l1 s1 q1 - l2 s12 (q1 + q2)
.
.
.
.
py = l1 c1 q1 + l2 c12 (q1 + q2)
.
.
.
$ = !z = q1 + q2

- l1 s1 - l2 s12
Jr(q) =

here, all rotations occur around the same

fixed axis z (normal to the plane of motion)

Robotics 1

py = l1 s1 + l2 s12

l1 c1 + l2 c12
1

- l2 s12
l2 c12
1

given r, this is a 3 x 2 matrix

12

Analytical Jacobian of polar robot

pz

direct kinematics (here, r = p)

q3

px = q3 c2 c1
py = q3 c2 s1

q2
d1

px

py
q1

fr(q)

pz = d1 + q3 s2

taking the time derivative

-q3c2s1 -q3s2c1 c2c1 .

.
.
v = p = q3c2c1 -q3s2s1 c2s1 q = Jr(q) q
0
q3c2 s2
+fr(q)
Robotics 1

+q

13

Geometric Jacobian
always a 6 x n matrix

JL(q)
JA(q)

.
q=

JL1(q) JLn(q)
J (q) J (q)
A1

An

.
q1

end-effector v
instantaneous E
!E
velocity

.
qn

superposition of effects
.
.
vE = JL1(q) q1 ++ JLn(q) qn

contribution to the linear

.
e-e velocity due to q1

.
.
!E = JA1(q) q1 ++ JAn(q) qn

contribution to the angular

.
e-e velocity due to q1

linear and angular velocity belong to

(linear) vector spaces in R3
Robotics 1

14

Contribution of a prismatic joint

note: joints beyond the i-th one are considered to be frozen,
so that the distal part of the robot is a single rigid body

.
.
JLi(q) qi = zi-1 di

zi-1

.
JLi(q) qi

qi = di
RF0

Robotics 1

joint i

.
JAi(q) qi

prismatic
i-th joint
.

zi-1 di
0

15

Contribution of a revolute joint

.
JLi(q) qi

.
.
JAi(q) qi = zi-1 ,i

zi-1

pi-1,E

Oi-1
qi =

RF0

Robotics 1

revolute
i-th joint

,i
joint i

.
JLi(q) qi
.
JAi(q) qi

(zi-1 ! pi-1,E) ,i
.

zi-1 ,i
16

Expression of geometric Jacobian

!E

=)

vE
!E

JL(q) .
q=
JA(q)

JL1(q) JLn(q)
J (q) J (q)

prismatic
i-th joint

zi-1 =

An

A1

revolute
i-th joint

JLi(q)

zi-1

zi-1 ! pi-1,E

JAi(q)

zi-1

0R (q )i-2R (q )
1 1
i-1 i-1

0
0
1

pi-1,E = p0,E(q1,,qn) - p0,i-1(q1,,qi-1)

Robotics 1

.
q1

.
p0,E

.
qn

this can be also

computed as

+p0,E
+qi

all vectors should be

expressed in the same
reference frame
(here, the base frame RF0)
17

Example: planar 2R arm

y2
y1
y0

z0 = z1 = z2 =

J=

Robotics 1

l2

l1
x0

DENAVIT-HARTENBERG table

x2

'i

di

ai

,i

l1

q1

l2

q2

x1
0
0
1

0A
1

z0 ! p0,E z1 ! p1,E
z0

joint

z1

0A
2

c1

- s1

l1c1

s1

c1

l1s1

c12

- s12

l1c1+ l2c12

s12

c12

l1s1+ l2s12

p0,1
p1,E = p0,E - p0,1

p0,E

18

Geometric Jacobian of planar 2R arm

y2
y1
y0

l1

l2

x2
J=

z0 ! p0,E z1 ! p1,E
z0

x1

x0
note: the Jacobian is here a 6!2 matrix,
thus its maximum rank is 2

- l1s1- l2s12
l1c1+ l2c12
0

at most 2 components of the linear/angular

end-effector velocity can be independently assigned
Robotics 1

z1

0
0
1

- l2s12
l2c12
0
0
0
1

compare rows 1, 2, and 6

with the analytical Jacobian
in slide #12!

19

Transformations of the Jacobian matrix

b) we may choose
E Oj(q)
On

RFi
Oj
0v

RF0

the one just

computed
n

0!
Bv

RFB

B!

0J (q)
n

=
=

a) we may choose
RFB RFi(q)
Robotics 1

rnE

.
q

vE = vn + ! ! rnE
= vn + S(rEn) !

BR
0

S(0rEn)

0v

BR
0

0!

BR (q)
0

S(0rEn(q))

BR (q)
0

0J (q)
n

. B
.
q = JE(q) q

never singular!
20

Example: Dexter robot

!

8R robot manipulator with transmissions by

pulleys and steel cables (joints 3 to 8)
!
!
!
!
!

Robotics 1

lightweight: only 15 kg in motion

motors located in second link
incremental encoders (homing)
redundancy degree for e-e pose task: n-m=2
compliant in the interaction with environment

21

Mid-frame Jacobian of Dexter robot

!
!

geometric Jacobian 0J8(q) is very complex

mid-frame Jacobian 4J4(q) is relatively simple!

08

y4
z4
6 rows,
8 columns

x0

x4
04

z0
y0

Robotics 1

22

Summary of differential relations

.
p.v
.
$.!

.
p=v
!=

!$. 1 +

!$. 2 +

!$. 3 =

.
.
.
.
a1 $1 + a2($1) $2 + a3($1, $2) $3 = T($)$

(moving) axes of definition for the sequence of rotations $i

r=

p
$

J(q) =

0 T($)

T($) has always /

a singularity
.
R.!
Robotics 1

.
R = S(!) R

Jr(q)

Jr(q) =

I
0

0
T-1($)

J(q)

singularity of the specific

minimal representation of orientation
for each column ri of R (unit vector of a frame),
.
we have

ri = ! ! ri

23

Acceleration relations (and beyond)

Higher-order differential kinematics

differential relations between motion in the joint space and motion in

the task space can be established at the second order, third order, ...
the analytical Jacobian always weights the highest-order derivative

velocity
acceleration

.
.
r = Jr(q) q

matrix function N2(q,q)

..
.
.. .
r = Jr(q) q + Jr(q) q

jerk

.
...
.. ..
.
...
r = Jr(q) q + 2 Jr(q) q + Jr(q) q

snap

.
.
r = Jr(q) q +

. ..

matrix function N3(q,q,q)

the same holds true also for the geometric Jacobian J(q)

Robotics 1

24

Primer on linear algebra

given a matrix J: m ! n (m rows, n columns)
!

rank 0(J) = max # of rows or columns that are linearly independent

!
!
!

range 2(J) = vector subspace generated by all possible linear

also called image of J
combinations of the columns of J
2(J)={v 3 Rm : # 4 3 Rn, v = J 4}
!

dim(5(J)) = n - 0(J)

also called null space of J

2(J) + 5(JT) = Rm e 2(JT) + 5(J) = Rn

!

dim(2(J)) = 0(J)

kernel 5(J) = vector subspace of all vectors 4 3 Rn such that J"4 = 0

!

0(J) 1 min(m,n) (if equality holds, J has full rank)

if m = n and J has full rank, J is non singular and the inverse J-1 exists
0(J) = dimension of the largest non singular square submatrix of J

sum of vector subspaces V1 + V2 = vector space where any element v can be

written as v = v1 + v2, with v1 3 V1 , v2 3 V2

all the above quantities/subspaces can be computed using, e.g., Matlab

Robotics 1

25

Robot Jacobian
decomposition in linear subspaces and duality
space of
joint velocities

space of
task (Cartesian)
velocities

J
0

5(J)

2(J)

2(JT) + 5(J) = Rn

2(J) + 5(JT) = Rm

2(JT)

5(JT)
0

0
space of
joint torques

Robotics 1

dual spaces

dual spaces

JT
(in a given configuration q)

space of
task (Cartesian)
forces

26

Mobility analysis
!

0(J) = 0(J(q)), 2(J) = 2(J(q)), 5(JT)= 5(JT(q)) are locally defined, i.e.,
they depend on the current configuration q
2(J(q)) = subspace of all generalized velocities (with linear and/or
angular components) that can be instantaneously realized by the robot
end-effector when varying the joint velocities at the configuration q
if J(q) has max rank (typically = m) in the configuration q, the robot
end-effector can be moved in any direction of the task space Rm
if 0(J(q)) < m, there exist directions in Rm along which the robot endeffector cannot move (instantaneously!)
!

these directions lie in 5(JT(q)), namely the complement of 2(J(q)) to the

task space Rm, which is of dimension m - 0(J(q))

when 5(J(q)) % {0}, there exist non-zero joint velocities that produce
zero end-effector velocity (self motions)
!

this always happens for m<n, i.e., when the robot is redundant for the task

Robotics 1

27

Kinematic singularities
!

configurations where the Jacobian loses rank

/ loss of instantaneous mobility of the robot end-effector
for m = n, they correspond to Cartesian poses at which the number of
solutions of the inverse kinematics problem differs from the generic case
in a singular configuration, we cannot find a joint velocity that realizes a
desired end-effector velocity in an arbitrary direction of the task space
close to a singularity, large joint velocities may be needed to realize
some (even small) velocity of the end-effector
finding and analyzing in advance all singularities of a robot helps in
avoiding them during trajectory planning and motion control
!
!

when m = n: find the configurations q such that det J(q) = 0

when m < n: find the configurations q such that all m!m minors of J are
singular (or, equivalently, such that det [J(q) JT(q)] = 0)

finding all singular configurations of a robot with a large number of joints,

or the actual distance from a singularity, is a hard computational task

Robotics 1

28

Singularities of planar 2R arm

py

l2
l1

p
q2

q1

px
x
analytical Jacobian
.
.
- l1s1- l2s12 - l2s12 .
p= lc+lc
l2c12 q = J(q) q
1 1
2 12
!
!

direct kinematics
px = l1 c1 + l2 c12
py = l1 s1 + l2 s12

det J(q) = l1l2s2

singularities: arm is stretched (q2 = 0) or folded (q2 = *)

singular configurations correspond here to Cartesian points on the
boundary of the workspace
in general, these singularities separate regions in the joint space with
distinct inverse kinematic solutions (e.g., elbow up or down)

Robotics 1

29

Singularities of polar (RRP) arm

pz

direct kinematics

q3

px = q3 c2 c1
py = q3 c2 s1

q2
d1
px

q1
det J(q) = q32 c2

.
p=

analytical Jacobian
-q3s1c2 -q3c1s2 c1c2
q3c1c2 -q3s1s2 s1c2
0
q3c2
s2

.
.
q = J(q) q

singularities
!
!

pz = d1 + q3 s2

py

E-E is along the z axis (q2 = */2): simple singularity rank J = 2

third link is fully retracted (q3 = 0): double singularity rank J drops to 1

all singular configurations correspond here to Cartesian points internal

to the workspace (supposing no limits for the prismatic joint)

Robotics 1

30

Singularities of robots
with spherical wrist
!
!

n = 6, last three joints are revolute and their axes intersect at a point
without loss of generality, we set O6 = W = center of spherical wrist
(i.e., choose d6 = 0 in the DH table)
J(q) =

J21

J22

since det J(q1,,q5) = det J11 " det J22 , there is a decoupling property
!
!

J11

det J11(q1,,q3) = 0 provides the arm singularities

det J22(q4, q5) = 0 provides the wrist singularities

being J22 = [z3 z4 z5] (in the geometric Jacobian), wrist singularities
correspond to when z3, z4 and z5 become linearly dependent vectors
6 when either q5 = 0 or q5 = */2

inversion of J is simpler (block triangular structure)

the determinant of J will never depend on q1: why?

Robotics 1

31