Robotics 2

Robust Trajectory Control

Prof. Alessandro De Luca

 Problem formulation
  given the real robot, modeled by

  assuming an estimated feedback linearization control

  we would like to design so as to obtain

  asymptotic stability of the closed-loop system
  the best possible trajectory tracking performance
  the linear feedback choice is not enough...

  questions:
  which should be the conditions on the estimates?
  can we guarantee stability/performance, based on known
bounds on the uncertainties?
Robotics 2 2
 Closed-loop equations -1

  under uncertain conditions (estimated ≠ real dynamic

coefficients), feedback linearization is only approximate
and the closed-loop equations are still nonlinear

where depends on the amount of uncertainty

Robotics 2 3
 Closed-loop equations -2

  closed-loop state equations are written as

  closed-loop error equations with respect to a desired

are rewritten as note that errors
are defined here
with opposite
signs w.r.t. usual

Robotics 2 4
 Solution approach

  add an external robust control term/loop

  based on computable bounds on the uncertainties
  based on the theory of guaranteed stability for
nonlinear uncertain system
  Lyapunov-based analysis
  a discontinuous control law will result
  difficult to implement because of chattering effects
  smoothed version with only uniformly ultimate
boundedness (u.u.b. stability) of the tracking error

Robotics 2 5
 Working assumptions
1.  bound on the desired trajectory

2.  bound on the estimate of the robot inertia matrix

with , holding for all configurations

3.  bound on the estimate of nonlinear dynamic terms

with a known function , bounded for all

  as a general rule, exploiting the model structure (e.g., its
linear parameterization) may lead to more stringent bounds

Robotics 2 6
 Bound on the inertia matrix
  assumption 2. can always be satisfied, knowing some upper and
lower bounds (that always exist due to the positive definiteness)
on the inverse of the inertia matrix

  it is then sufficient to choose as estimate

with

  in fact, using the SVD factorization of the inverse inertia matrix, it

can be shown that (see Appendix A)

Robotics 2 7
 Control design – step 1
  linear control law with an added robust term

where the PD gains are diagonal and positive matrices

  we obtain

being

where has all eigenvalues with negative real part,

and

Robotics 2 8
 Control design – step 2
  (same) bound on nonlinear terms and added robust term

  we can use the previous data and implicitly define the

bound from

yielding the well-defined (since ), limited and

possibly time-varying function

Robotics 2 9
 Control design – step 3
  solve an associated (linear) Lyapunov equation, for any
given symmetric matrix

finding the unique (symmetric) solution matrix

  finally, define the discontinuous robust term as

that also satisfies, by its own structure,

Robotics 2 10
 Solving the Lyapunov equation
  in general, using lyap in Matlab (only once, in advance)
  closed-form solution in an interesting scalar case (one
robot joint/link), to get a “feeling”...

choose, e.g., ⇒ find

so that (also in the n-dof case)

Robotics 2 11
 Stability analysis
Theorem 1
Defining , the presented robust control law
with the discontinuous term is such that along
the trajectories of the closed-loop error system
Proof

if ⇒

if ⇒ ⇒
note: because of the discontinuity
we cannot directly conclude on the
(global) asymptotic stability of e=0 ⇒
Robotics 2 12
 A smoother robust controller

  for any given (small) , define the continuous

robust term as

Theorem 2
With the above continuous robust control law, any solution
, with , of the closed-loop error system is
uniformly ultimately bounded with respect to a suitable set
(a neighborhood of the origin)
Proof in Appendix B
Robotics 2 13
 Case study: Single-link under gravity
model (no friction)

error equations

known bounds for control design

Robotics 2 14
 Calculations for robust control
% real robot % bounding dynamic terms
I=5; mgd=7; % inertia
% initial robot state m=1/I_max; M=1/I_min;
th0=0;thp0=0; c=(M+m)/2;
% range of uncertainties alpha=(M-m)/(M+m);
I_min=5; I_max=10; Ihat=1/c; % =6.6667
mgd_min=5;mgd_max=7; % nonlinear terms (only gravity)
% linear tracking stabilizer gains Mphi=M*(mgd_max-mgd_min);
kp=25; kd=10; % two poles in -5 mgdhat=5;
% overall bounding
% robust control part rho0=Mphi/(1-alpha) % =0.6
% Lyapunov matrix P and b^T P term rho1=alpha/(1-alpha) % =0.5
A=[0 1; -kp -kd]; % smoothed version
q=1; Q=q*eye(2); epsilon=5*10^-4;
P=lyap(A',Q); % solve A'*P+P*A+Q=0
b=[0 1];
bP=b*P; % =[0.02 0.052] red values are used in Simulink

Robotics 2 15
 Results
first trajectory – feedback linearization, no robust loop
0.04 10

0.02
5

0
0
position error (rad)

torque (Nm)
−0.02
−5
−0.04

−10
−0.06

−0.08 position error −15

control torques
−0.1 −20
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time (s) time (s)

0.2

0.15

0.1
velocity error (rad/s)

0.05

non-zero initial error

0

on velocity
−0.05

−0.1

−0.15
velocity error
−0.2
0 5 10 15 20 25 30
time (s)

Robotics 2 16
 Results
first trajectory – discontinuous robust control
0.04 10

0.02
5

0
0
position error (rad)

torque (Nm)
−0.02
−5
−0.04

−10
−0.06

−0.08 position error −15

control torques
−0.1 −20
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time (s) time (s)

0.2

0.15

0.1
velocity error (rad/s)

position and velocity errors

0.05

are largely reduced,

0

−0.05

−0.1
but control chattering
−0.15
velocity error at high frequency
−0.2
0 5 10 15 20 25 30
(when error is close to zero)
time (s)

Robotics 2 17
 Results
first trajectory – smoothed robust control
0.04 10

0.02
5

0
0
position error (rad)

torque (Nm)
−0.02
−5
−0.04

−10
−0.06

−0.08 position error −15

control torques
−0.1 −20
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time (s) time (s)

0.2

0.15

0.1
velocity error (rad/s)

position and velocity errors

0.05

are similarly reduced,

0

−0.05

−0.1
without control chattering
−0.15
velocity error
−0.2
(using here )
0 5 10 15 20 25 30
time (s)

Robotics 2 18
 Results
second trajectory – fbk linearization, no robust loop
0.03 20

15
0.02

10

0.01
position error (rad)

torque (Nm)
0 0

−5
−0.01

−10

−0.02

position error −15

control torques
−0.03 −20
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time (s) time (s)

bang-bang acceleration profile

0.1

0.08

0.06 at 1 rad/s frequency and

0.04
with Qmax = 1 rad/s2
velocity error (rad/s)

0.02

−0.02 zero initial tracking error

−0.04
(matching state conditions)
−0.06

−0.08 velocity error

−0.1
0 5 10 15 20 25 30
time (s)

Robotics 2 19
 Results
second trajectory – discontinuous robust control
0.03 20

15
0.02

10

0.01
position error (rad)

torque (Nm)
0 0

−5
−0.01

−10

position error control torques

−0.02
−15

−0.03 −20
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time (s) time (s)

0.1

0.08
bang-bang acceleration profile
0.06

0.04

position and velocity errors

velocity error (rad/s)

0.02

again largely reduced,

0

−0.02

−0.04 but control chattering

−0.06

−0.08 velocity error and larger effort

−0.1
0 5 10 15 20 25 30
time (s)

Robotics 2 20
 Results
second trajectory – smoothed robust control
0.03 20

15
0.02

10

0.01
position error (rad)

torque (Nm)
0 0

−5
−0.01

−10

position error control torques

−0.02
−15

−0.03 −20
0 5 10 15 20 25 30 0 5 10 15 20 25 30
time (s) time (s)

0.1

0.08 bang-bang acceleration profile

0.06

0.04 position and velocity errors

velocity error (rad/s)

0.02
are further reduced,
without control chattering
0

−0.02

−0.04
and same control effort as
velocity error without robustifying term
−0.06

−0.08

−0.1
0 5 10 15
time (s)
20 25 30 (using here )
Robotics 2 21
 Appendix A
Proof of bounds on the inertia matrix

  the SVD factorization of the (symmetric) inverse inertia matrix is

so that, with the choice made for its estimate, it follows that

Robotics 2 22
 Appendix B
Proof of Theorem 2

  setting , note that for the robust term in the control

law it is

  defining as before , we have

having used the chain of inequalities

  if , the rest of the proof is the same as in Theorem 1

%
Robotics 2 23
 Appendix B
Proof of Theorem 2 (cont)

  if , the second term in the derivative of is

with a maximum value attained for

  therefore, it is

provided that

  if is the smallest level set of (an ellipsoid) containing

the hyper-sphere of radius , then
and u.u.b. is obtained for
(an upper bound for the time needed to reach can be given)
Robotics 2 24