5. Going into the real world
Unlike industrial settings, the world
in which we live changes all the
time:
Cannot predict all circumstances
Need to react rapidly and
appropriately
6. Robots that deal with uncertainty
Commercial airplanes fly autonomously to a very large extent.
They can recover from turbulences automatically.
7. Robots that deal with uncertainty
Automobile industry and governments support research
to build fully autonomous cars.
8. Uncertainty in home environment
No two apartments
look the same
And the same apartment
can change appearance
from one day to the next
9. Variability in task definition
What does it mean to grate carrots?
More than one way to do this.
More than one tool to
perform the task.
11. Learning from Human Demonstrations
Learning a skill is more than simply replaying a trajectory.
It requires to understand what a skill is.
To learn this, one needs to show several demonstrations
to generalize across sets of examples.
12. Learning from Human Demonstrations
Learning a skill is more than simply replaying a trajectory.
It requires to understand what a skill is.
To learn this, one needs to show several demonstrations
to generalize across sets of examples.
K. Kronander, M. Khansari and A. Billard, JTSC Best Paper Award, Int. Conf. on Intelligent and Robotics Systems, IROS 2011
15. Teaching robots to be less stiff
Being stiff is not always good How to teach a robot to relax…
Low stiffness when carrying the liquid High stiffness when pouring the liquid
http://lasa.epfl.ch
16. Teaching robots to be less stiff
Being stiff is not always good How to teach a robot to relax…
Shaking the robot: A natural method to teach a robot to relax.
http://lasa.epfl.ch
17. Teaching robots to be less stiff
After training the robot manages to adapt naturally when
required and remains stiff when required.
http://lasa.epfl.ch
19. Learning from Failure
The robot is provided solely with failed examples.
It has no information about the task – no reward, no indication of
what was incorrect.
Training examples Reproduction
D. Grollman and A. Billard, Best Cognitive Robotics Paper Award, Int. Conf. on Robotics and Automation, ICRA 2011
20. Learning from Failure
Find a solution in a few trials
Is comparable in efficiency to classical reinforcement learning approaches
But does not need a reward!
22. Generalizing: Learning a control law
Learning a control law that ensures that you reach the target even if perturbed
and that you follow a particular dynamics
http://lasa.epfl.ch
23. Coupled Dynamical Systems
Coupled control of hand andand fingers may lead fingers and
Decoupled control of hand fingers ensures that to
hand close in a coordinated manner perturbations.
failure when adapting to very rapid on the new target.
http://lasa.epfl.ch
24. CoupledCoupled Dynamical Systems Grasp
Dynamical Systems for Reach and
Adaptation to perturbation of the order of a few millisecunds.
http://lasa.epfl.ch
26. Catching Objects in Flight
Extremely fast computation (object flies in half a second); re-estimation of
arm motion to adapt to noisy visual detection of object.
http://lasa.epfl.ch
27. Catching Objects in Flight
STEP 1: Build a model of the graspable region on the object;
http://lasa.epfl.ch
28. Catching Objects in Flight
STEP 1: Build a model of the graspable region on the object; learn likelihood
of placing fingers in region of the handle from several demonstrations;
z
x
0.04
0.03
0.02
0.01
Z (m)
Z (m)
0
-0.01
-0.02
-0.03
-0.04
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
X (m)
X (m)
http://lasa.epfl.ch
29. Catching Objects in Flight
STEP 1: Build a model of the graspable region on the object; learn likelihood
of placing fingers in region of the handle from several demonstrations;
z
x
y
z
x
y
0.04 0.04
0.03 0.03
0.02 0.02
0.01 0.01
Y (m)
Z (m)
Z (m)
Y (m)
z 0
-0.01
0
-0.01
-0.02 -0.02
-0.03 -0.03
-0.04
-0.04
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 -0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
X (m) X (m)
x
X (m) X (m)
http://lasa.epfl.ch
30. Catching Objects in Flight
STEP2: Learn a model of the translational and rotational motion of the object
If the motion of the object is complex, a simple
ballistic model is not sufficient
Needs to estimate the dynamics of object in flight http://lasa.epfl.ch
31. Catching Objects in Flight
STEP2: Gather several examples
Use non-linear regression model (Support
Vector Regression)
Precision (1cm, 1degree); computation
0.17-0.32 second ahead of time.
Combine with extended Kalman Filter to
tackle innacuracy of vision.
http://lasa.epfl.ch
S. Kim and A. Billard, Aut. Robots, 2012
32. Catching Objects in Flight
STEP 3: Compute the region of feasible
hand postures that yield a possible grasp.
http://lasa.epfl.ch
33. Catching Objects in Flight
STEP 3: Compute the region of feasible hand postures that yield a possible
grasp through sampling space.
Probability Contour
0.5
3
0.4
2
0.3
0.2 1
0.1
Y (m)
sY
0
0
-1
-0.1
-0.2 -2
-0.3
-3
-0.4
-0.6 -0.4 -0.2 0 0.2 0.4 -5 -4 -3 -2 -1 0 1 2 3 4
X (m) sX
0.5
0.4 3
0.3
2
0.2
1
Z (m)
0.1
sZ
0 0
-0.1
-1
-0.2
-2
-0.3
-0.4 -3
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4
X (m)
-5 -4 -3 -2 -1 0 1 2 3 4
sX
3 0.45
0.5 1
0.8
2 0.4
0.4
0.6
0.4 1 0.35
0.3
Z (m)
0.2
0.3
sZ
0
sZ
0.2 0
-0.2 0.25
-1
0.1 -0.4
0.2
-0.6 -2
0
-0.8 0.15
-3
-1
0.1
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6
-6 -5 -4 -3 -2 -1 0 1 2 3
Y (m) -1.5 -1 -0.5 0 0.5 1 1.5
sY
Position Orientation sX
http://lasa.epfl.ch
34. Catching Objects in Flight
STEP 4: Find the grasping posture by
predicting dynamics of motion and finding
most likely combination of grasping point
and feasible hand posture.
http://lasa.epfl.ch
35. Catching Objects in Flight
STEP 5: Generate motion of hand and fingers
to catch the object at the right place using
coupled dynamical systems for hand position
and orientation and for finger motion.
http://lasa.epfl.ch
36. Catching Objects inobject
Catching a flying Flight
Kim, Shukla and Billard: In preparation http://lasa.epfl.ch
37. Catching Objects inobject
Catching a flying Flight
Kim, Shukla and Billard: In preparation http://lasa.epfl.ch