Robotics 1

Inverse kinematics
Prof. Alessandro De Luca

Robotics 1 1
 Inverse kinematics
what are we looking for?

direct kinematics is always unique;

how about inverse kinematics for this 6R robot?
Robotics 1 2
 Inverse kinematics problem
n “given a desired end-effector pose (position +
orientation), find the values of the joint variables
that will realize it”
n a synthesis problem, with input data in the form
R p p
n T= = 0A
n(q) §r= = fr(q), or for any
000 1 f other task vector
classical formulation: generalized formulation:
inverse kinematics for a given end-effector pose inverse kinematics for a given value of task variables

n a typical nonlinear problem

n existence of a solution (workspace definition)
n uniqueness/multiplicity of solutions (r Î Rm, q ÎRn)
n solution methods
Robotics 1 3
 Solvability and robot workspace
(for tasks related to a desired end-effector Cartesian pose)

n primary workspace WS1: set of all positions p that can be

reached with at least one orientation (f or R)
n out of WS1 there is no solution to the problem
n when p Î WS1, there is a suitable f (or R) for which a solution exists

n secondary (or dexterous) workspace WS2: set of positions p

that can be reached with any orientation (among those
feasible for the robot direct kinematics)
n when p Î WS2, there exists a solution for any feasible f (or R)

n WS2 Í WS1

Robotics 1 4
 Workspace of Fanuc R-2000i/165F
section for a
constant angle q1
WS1⊂R3
(≈ WS2 for spherical wrist
without joint limits)

rotating the
base joint angle q1
Robotics 1 5
 Workspace of planar 2R arm
2 orientations
if p Î int(WS1) l1+l2
y • p = WS1- ¶WS1
l2
q2
|l1-l2|
l1
q1 ¶WS1
outer and inner
x boundaries 1 orientation
n if l1 ¹ l2
n WS1 = {p Î R2: |l1-l2| £ ║p║£ l1+l2} ⊂ R2
n WS2 = Æ
n if l1 = l2 = ℓ
n WS1 = {p Î R2: ║p║£ 2ℓ} ⊂ R2
n WS2 = {p = 0} (infinite number of feasible orientations at the origin)

Robotics 1 6
 Wrist position and E-E pose
inverse solutions for an articulated 6R robot
LEFT DOWN Unimation PUMA 560 RIGHT DOWN

4 inverse solutions
out of singularities
(for the position of
the wrist center only)

LEFT UP 8 inverse solutions considering RIGHT UP

the complete E-E pose
(spherical wrist: 2 alternative
solutions for the last 3 joints)

Robotics 1 7
 Inverse kinematic solutions of UR10
6-dof Universal Robot UR10, with non-spherical wrist
video (slow motion)

desired pose
−0.2373
p = −0.0832 [m]
1.3224
3/2 0.5 0
R = −0.5 3/2 0
0 0 1
home configuration at start
0 = 0 −1/2 0 −1/2 0 0 T

[rad]

Robotics 1 9
 The 8 inverse kinematic solutions of UR10

shoulderRight shoulderRight shoulderRight shoulderRight

wristDown wristDown wristUp wristUp
elbowUp elbowDown elbowUp elbowDown
1.0472 1.0472 1.0472 1.0472
−1.2833 −1.9941 −1.5894 −2.0944
−0.7376 0.7376 −0.5236 0.5236
!= != != !=
−2.6915 2.8273 0.5422 0
−1.5708 −1.5708 1.5708 1.5708
3.1416 3.1416 0 0

shoulderLeft shoulderLeft shoulderLeft shoulderLeft

wristDown wristDown wristUp wristUp
elbowDown elbowUp elbowDown elbowUp
2.7686 2.7686 2.7686 2.7686
−1.0472 −1.5522 −1.1475 −1.8583
−0.5236 0.5236 −0.7376 0.7376
!= != != !=
3.1416 2.5994 0.3143 −0.4501
−1.5708 −1.5708 1.5708 1.5708
1.4202 1.4202 −1.7214 −1.7214

Robotics 1 10
 Multiplicity of solutions
some examples

n E-E positioning (m=2) of a planar 2R robot arm

n 2 regular solutions in int(WS1)
n 1 solution on ¶WS1
singular solutions
n for l1 = l2: ¥ solutions in WS2
n E-E positioning of an articulated elbow-type 3R robot arm
n 4 regular solutions in WS1 (with singular cases yet to be investigated ...)
n spatial 6R robot arms
n £ 16 distinct solutions, out of singularities: this “upper bound” of
solutions was shown to be attained by a particular instance of
“orthogonal” robot, i.e., with twist angles ai = 0 or ±p/2 ("i)
n analysis based on algebraic transformations of robot kinematics
n transcendental equations are transformed into a single polynomial
equation of one variable
n seek for an equivalent polynomial equation of the least possible degree
Robotics 1 11
 A planar 3R arm
workspace and number/type of inverse solutions
y q3 l1 = l2 = l3 = ℓ, n=3, m=2
• p
l3 WS1 = {p Î R2: ║p║£ 3ℓ} ⊂ R2
l2
q2
WS2 = {p Î R2: ║p║£ ℓ} ⊂ R2
l1
q1 any planar orientation is feasible in WS2
x
1. in WS1 : ¥1 regular solutions (except for 2. and 3.),
at which the E-E can take a continuum of
¥ orientations (but not all orientations in the plane!)
2. if ║p║= 3ℓ : only 1 solution, singular
3. if ║p║= ℓ : ¥1 solutions, 3 of which singular

4. if ║p║< ℓ : ¥1 regular solutions (never singular)

Robotics 1 12
 Workspace of a planar 3R arm
general case: different link lengths

⇒ lmin= l3 = 0.3

⇒ 0.5
Rin= 0,
Robotics 1 13
 Multiplicity of solutions
summary of the general cases

n if m = n
n ∄ solutions
n a finite number of solutions (regular/generic case)
n “degenerate” solutions: infinite or finite set, but anyway
different in number from the generic case (singularity)
n if m < n (robot is redundant for the kinematic task)
n ∄ solutions
n ¥n-m solutions (regular/generic case)
n a finite or infinite number of singular solutions
n use of the term singularity will become clearer when dealing
with differential kinematics
n instantaneous velocity mapping from joint to task velocity
n lack of full rank of the associated m n Jacobian matrix J(q)
Robotics 1 14
 Dexter robot (8R arm)
n m = 6 (position and orientation of E-E)
n n = 8 (all revolute joints)
n ¥2 inverse kinematic solutions (redundancy degree = n-m = 2)

video

exploring inverse kinematic solutions by a self-motion

Robotics 1 15
 Solution methods

ANALYTICAL solution NUMERICAL solution

(in closed form) (in iterative form)

§ preferred, if it can be found* § certainly needed if n>m (redundant

§ use ad-hoc geometric inspection case), or at/close to singularities
§ algebraic methods (solution of § slower, but easier to be set up
polynomial equations) § in its basic form, it uses the
§ systematic ways for generating a (analytical) Jacobian matrix of the
reduced set of equations to be direct kinematics map
solved ¶fr (q)
* sufficient conditions for 6-dof arms
Jr(q) =
¶q
• 3 consecutive rotational joint axes are
incident (e.g., spherical wrist), or
§ Newton method, Gradient method,
• 3 consecutive rotational joint axes are and so on…
parallel
D. Pieper, PhD thesis, Stanford University, 1968
Robotics 1 16
 Inverse kinematics of planar 2R arm
y
py • p direct kinematics
l2 px = l1 c1 + l2 c12
q2
l1 py = l1 s1 + l2 s12
q1
data q1, q2 unknowns
px x

“squaring and summing” the equations of the direct kinematics

px2 + py2 - (l12 + l22) = 2 l1 l2 (c1 c12 + s1 s12) = 2 l1 l2 c2
in analytical form
and from this

c2 = (px2 + py2 - l12 - l22)/ 2 l1 l2, s2 = ±Ö1 - c22 q2 = ATAN2 {s2, c2}

must be in [-1,1] (else, point p 2 solutions

is outside robot workspace!)
Robotics 1 17
 Inverse kinematics of 2R arm (cont’d)
y
py • p
by geometric inspection
q2
q1 = a - b
b

q1 a
px x
2 solutions
q1 = ATAN2 {py, px} - ATAN2 {l2 s2 , l1 + l2 c2}
(one for each value of s2)
note: difference of ATAN2 needs
to be re-expressed in (-p , p]!
q2’’
• p
{q1,q2}UP/LEFT q2’ {q1,q2}DOWN/RIGHT

q1” q2’ e q2’’ have same absolute

value, but opposite signs
q1’
Robotics 1 18
 Algebraic solution for q1

another px = l1 c1 + l2 c12 = l1 c1 + l2 (c1 c2 - s1 s2)

linear in
solution
py = l1 s1 + l2 s12 = l1 s1 + l2 (s1c2 + c1s2) s1 and c1
method…

l1 + l2c2 - l2s2 c1 px
=
l2s2 l1 + l2c2 s1 py

except for l1=l2 and c2=-1

det = (l12 + l22 + 2 l1l2c2) > 0 being then q1 undefined
(singular case: ¥1 solutions)
q1 = ATAN2 {s1, c1} = ATAN2 {(py(l1+l2c2)-pxl2s2)/det, (px(l1+l2c2)+pyl2s2)/det}

notes: a) this method provides directly the result in (-p , p]

b) when evaluating ATAN2, det > 0 can be eliminated
from the expressions of s1 and c1
Robotics 1 19
 Inverse kinematics of polar (RRP) arm
pz
Note: px = q3 c2 c1
q2 is NOT a q3
DH variable!
py = q3 c2 s1
pz = d1 + q3 s2
q2
d1 px2 + py2 + (pz - d1)2 = q32
py
q3 = + Öpx2 + py2 + (pz - d1)2
px q1 our choice: take here only the positive value...

if q3 = 0, then q1 and q2 remain both undefined (stop); else

(if it stops,
q2 = ATAN2{(pz - d1)/q3, ± Ö(px2 + py2)/q32 } a singular case:
¥2 or ¥1
if px2+py2 = 0, then q1 remains undefined (stop); else solutions)
q1 = ATAN2{py/c2, px/c2} (2 regular solutions {q1,q2,q3})
Robotics 1 we have eliminated q3>0 from both arguments! 20
 Inverse kinematics of 3R elbow-type arm
{u, d} = elbow up, down
L2

q3
q2
L3
d1 pz
{f, b} = facing, backing
the point p=(px, py, pz) py

q1 symmetric structure without offsets

px e.g., first 3 joints of Mitsubishi PA10 robot

px = c1 (L2c2+ L3c23)
direct four regular inverse
py = s1 (L2c2+ L3c23)
kinematics kinematics solutions in WS1
pz = d1+ L2s2+ L3s23
Note: more details (e.g., full handling of
WS1={spherical shell centered at (0,0,d1), with outer singular cases) can be found in the solution
radius Rout= L2 +L3 and inner radius Rin=|L2-L3|} of the Robotics 1 written exam of 11.04.2017
Robotics 1 21
 Inverse kinematics of 3R elbow-type arm
L2
q3 px = c1 (L2c2+ L3c23)
q2 py = s1 (L2c2+ L3c23) direct
L3 kinematics
d1 pz pz = d1+ L2s2+ L3s23
py

q1
px

px2 + py2 + (pz-d1)2 = c12 (L2c2+ L3c23)2 + s12 (L2c2+ L3c23)2 + (L2s2+ L3s23)2
= ... = L22 + L32 + 2L2L3 (c2c23+s2s23) = L22 + L32 + 2L2L3 c3

c3 = (px2 + py2 + (pz-d1)2 - L22 - L32) / 2L2L3 ∈ [-1,1] (else, p is out of workspace!)

q3{+} = ATAN2{s3, c3}

s3 = Ö1 - c32 two solutions
q3{-} = ATAN2{-s3, c3} = - q3{+}
Robotics 1 22
 Inverse kinematics of 3R elbow-type arm
L2
q3 px = c1 (L2c2+ L3c23)
q2 py = s1 (L2c2+ L3c23) direct
L3 kinematics
d1 pz pz = d1+ L2s2+ L3s23
py

q1
px (being px2 + py2 = (L2c2+ L3c23)2 > 0)

only when px2 + py2 >0 c1 = px / √px2 + py2

(else q1 is undefined —infinite solutions!) s1 = py / √px2 + py2

q1{+} = ATAN2{py, px}

again, two solutions
q1{-} = ATAN2{-py, -px}

Robotics 1 23
 Inverse kinematics of 3R elbow-type arm
combine the first two direct kinematics
L2
equations and rearrange the last one
q3 c1px + s1py = L2c2 + L3c23
q2
L3 = (L2+L3c3) c2 – L3s3 s2
d1 pz pz – d1 = L2s2+ L3s23
py = L3s3 c2 + (L2+L3c3) s2

define and solve a linear system Ax = b

q1
px in the algebraic unknowns x = (c2, s2)

four regular solutions for q2,

depending on combinations
of {+,-} from q1 and q3
coefficient matrix A known vector b

provided det A = px2 + py2 + (pz-d1)2 > 0

(else q2 is undefined —infinite solutions!) q2{{f,b},{u,d}}
= ATAN2{s2{{f,b},{u,d}}, c2{{f,b},{u,d}}}
Robotics 1 24
 Inverse kinematics
for robots with spherical wrist
z4 y6
j6 W x6
first 3 robot joints O6 = p
of any type (RRR, RRP, PPP, …)
z3 z5 z6 = a

z0 j5 d6
j4

y0 find q1, …, q6 from the input data:

• p (origin O6)
x j1 • R = [n s a] (orientation of RF6)
0
1. W = p - d6 a ® q1, q2, q3 (inverse “position” kinematics for main axes)
2. R = 0R3(q1, q2, q3) 3R6(q4, q5, q6) ® 3R6(q4, q5, q6) = 0R3T R ® q4, q5, q6
(inverse “orientation”
given known, Euler ZYZ or ZXZ kinematics for the wrist)
after step 1 rotation matrix
Robotics 1 25
 6R example: Unimation PUMA 600

spherical n = 0x6(q)
wrist

s = 0y6(q)

a function of
q1, q2, q3 only!
a = 0z6(q)

p = 06(q)

8 different inverse solutions

here d6=0, that can be found in closed form
so that 06=W directly (see Paul, Shimano, Mayer; 1981)
Robotics 1 26
 Numerical solution of
inverse kinematics problems
n use when a closed-form solution q to rd = fr(q) does not exist
or is “too hard” to be found
¶fr (analytical Jacobian)
n Jr(q) =
¶q
n Newton method (here for m=n)
n rd = fr(q) = fr(qk) + Jr(qk) (q - qk) + o(║q - qk║2) ¬ neglected

qk+1 = qk + Jr-1(qk) [rd - fr(qk)]

n convergence if q0 (initial guess) is close enough to some q*: fr(q*) = rd

n problems near singularities of the Jacobian matrix Jr(q)
n in case of robot redundancy (m<n), use the pseudo-inverse Jr#(q)
n has quadratic convergence rate when near to solution (fast!)

Robotics 1 27
 Operation of Newton method
n in the scalar case, also known as “method of the tangent”
n for a differentiable function f(x), find a root of f(x*)=0 by iterating as

an approximating sequence

animation from
http://en.wikipedia.org/wiki/File:NewtonIteration_Ani.gif

Robotics 1 28
 Numerical solution of
inverse kinematics problems (cont’d)

n Gradient method (max descent)

n minimize the error function
H(q) = ½ ║rd - fr(q)║2 = ½ [rd - fr(q)]T [rd - fr(q)]
qk+1 = qk - a ÑqH(qk)
from ÑqH(q) = - JrT(q) [rd - fr(q)], we get
qk+1 = qk + a JrT(qk) [rd - fr(qk)]

n the scalar step size a > 0 should be chosen so as to guarantee

a decrease of the error function at each iteration (too large
values for a may lead the method to “miss” the minimum)
n when the step size a is too small, convergence is extremely slow

Robotics 1 29
 Revisited as a “feedback” scheme
q(0)
.
rd + e q ó q rd = cost
JrT(q) õ
-
r (a = 1)
fr(q)

e = rd - fr(q) ® 0 Û closed-loop equilibrium e=0 is asymptotically stable

V = ½ eTe ³ 0 Lyapunov candidate function
. . .
V=e e =e
T T d ((rd - fr(q)) = - e Jr q = - eT Jr JrTe £ 0
T

. dt
V = 0 Û e Î Ker(JrT) in particular e = 0

asymptotic stability
Robotics 1 30
 Properties of Gradient method
n computationally simpler: Jacobian transpose, rather than its
(pseudo)-inverse
n direct use also for robots that are redundant for the task
n may not converge to a solution, but it never diverges
n the discrete-time evolution of the continuous scheme

qk+1 = qk + DT JrT(qk) [rd - f(qk)] (a = DT)

is equivalent to an iteration of the Gradient method
n scheme can be accelerated by using a gain matrix K>0
.
q = JrT(q) K e
note: K can be used also to “escape” from being stuck in a stationary point,
by rotating the error e out of the kernel of JrT (if a singularity is encountered)

Robotics 1 31
 A case study
analytic expressions of Newton and gradient iterations

n 2R robot with l1 = l2 = 1, desired end-effector position rd = pd = (1,1)

n direct kinematic function and error

n Jacobian matrix

n Newton versus Gradient iteration

det Jr(q)
ek

Robotics 1 32
 Error function
n 2R robot with l1=l2=1, desired end-effector position pd = (1,1)

e = pd - fr(q)

2
plot of ║e║ as a function of q = (q1,q2) two local minima
(inverse kinematic solutions)
Robotics 1 33
 Error reduction by Gradient method
n flow of iterations along the negative (or anti-) gradient
n two possible cases: convergence or stuck (at zero gradient)

start

one solution

.
local maximum
.
saddle point
(stop if this is the initial guess) (stop after some iterations)

another start...

...the other solution

(q1,q2)’ =(0,π/2) (q1,q2)” =(π/2,-π/2) (q1,q2)max =(-3π/4,0) (q1,q2)saddle =(π/4,0)

Robotics 1 e Î Ker(JrT) ! 34
 Convergence analysis
when does the gradient method get stuck?

n lack of convergence occurs when

n the Jacobian matrix Jr(q) is singular (the robot is in a “singular configuration”)
n AND the error is in the “null space” of JrT(q)
p

pd
q2

q1
pd

(q1,q2)saddle =(π/4,0) e Î Ker(JrT)

pd e ∉ Ker(JrT) !!
e Î Ker(JrT) the algorithm will
proceed in this case,
moving out of
p the singularity
p (q1,q2)max =(-3π/4,0) (q1,q2) =(π/9,0)
Robotics 1 35
 Issues in implementation
n initial guess q0
n only one inverse solution is generated for each guess
n multiple initializations for obtaining other solutions
n optimal step size a in Gradient method
n a constant step may work good initially, but not close to the
solution (or vice versa)
n an adaptive one-dimensional line search (e.g., Armijo’s rule) could
be used to choose the best a at each iteration
n stopping criteria
Cartesian error algorithm
(possibly, separate for ║rd - f(qk)║ ≤ ε ║qk+1-qk║ ≤ εq
increment
position and orientation)
n understanding closeness to singularities
numerical conditioning
σmin {J(qk)} ≥ σ0 of Jacobian matrix (SVD)
(or a simpler test on its determinant, for m=n)

Robotics 1 36
 Numerical tests on RRP robot
n RRP/polar robot: desired E-E position rd = pd = (1, 1 ,1)
—see slide 20, with d1=0.5
n the two (known) analytical solutions, with q3 ≥ 0, are:
q* = (0.7854, 0.3398, 1.5)
q** = (q1*- p, p - q2*, q3*) = (-2.3562, 2.8018, 1.5)
n norms ε = 10-5 (max Cartesian error), εq =10-6 (min joint increment)
n kmax=15 (max # iterations), |det(Jr)| ≤ 10-4 (closeness to singularity)
n numerical performance of Gradient (with different steps a) vs. Newton
n test 1: q0 = (0, 0, 1) as initial guess
n test 2: q0 = (-p/4, p/2, 1) —”singular” start, since c2=0 (see slide 20)
n test 3: q0 = (0, p/2, 0) —”double singular” start, since also q3=0
n solution and plots with Matlab code

Robotics 1 37
 Numerical test -1
n test 1: q0 = (0, 0, 1) as initial guess; evolution of error norm

Gradient: a = 0.5 Gradient: a = 1 Gradient: a = 0.7

too large, oscillates

around solution good, converges
in 11 iterations
slow, 15 (max)
iterations

0.57⋅10-5
Newton
ex
Cartesian errors
component-wise
very fast, converges
in 5 iterations ey

ez

Robotics 1 0.15⋅10-8 38
 Numerical test -1

n test 1: q0 = (0, 0, 1) as initial guess; evolution of joint variables

Gradient: a=1 Gradient: a = 0.7 Newton

not converging converges in converges in
to a solution 11 iterations 5 iterations

both to the same solution q* = (0.7854, 0.3398, 1.5)

Robotics 1 39
 Numerical test -2
n test 2: q0 = (-p/4, p/2, 1): singular start
with check of
Newton singularity:
blocked at start
error norms

without check:
it diverges!
Gradient
a = 0.7

!!
starts toward
solution, but
joint variables

slowly stops
(in singularity):
when Cartesian error
vector e ∈ Ker(JrT)

Robotics 1 40
 Numerical test -3
n test 3: q0 = (0, p/2, 0): “double” singular start

Gradient (with a = 0.7) Newton

starts toward solution is either
error norm

exits the double singularity blocked at start

slowly converges in 19 or (w/o check)
0.96⋅10-5 iterations to the solution explodes!
q =(0.7854, 0.3398, 1.5)
*
→ “NaN” in Matlab
Cartesian errors

joint variables

Robotics 1 41
 Final remarks
n an efficient iterative scheme can be devised by combining
n initial iterations using Gradient (“sure but slow”, linear convergence rate)
n switch then to Newton method (quadratic terminal convergence rate)
n joint range limits are considered only at the end
n check if the solution found is feasible, as for analytical methods
n in alternative, an optimization criterion can be included in the search
n driving iterations toward an inverse kinematic solution with nicer properties
n if the problem has to be solved on-line
n execute iterations and associate an actual robot motion: repeat steps at
times t0, t1=t0+T, ..., tk=tk-1+T (e.g., every T=40 ms)
n the “good” choice for the initial guess q0 at tk is the solution of the previous
problem at tk-1 (provides continuity, needs only 1-2 Newton iterations)
n crossing of singularities/handling of joint range limits need special care
n Jacobian-based inversion schemes are used also for kinematic control,
along a continuous task trajectory rd(t)
Robotics 1 42