Dexterous Robotics Laboratory, Johnson Space Center

1
Grasp Synthesis by Sequencing Contact-Relative
Motions
Robert Platt Dexterous Robotics Laboratory,
Robotics, Automation, and Simulation Division,
Johnson Space Center, NASA
Dexterous Robotics Laboratory, Johnson Space
Center
2
Two Extremes

• Force-Moment Residual Grasp Control (Coelho,
  Platt, Grupen)
• little prior knowledge needed (only assumption
  object must be convex)
• slower execution the need to maintain contact
  with the object can slow down contact
  displacements
• 2. Geometry Based Grasp Planning (Faverjohn,
  Ponce, Sudsang, Trinkle, many others)
• requires complete object geometry and pose
• fast execution position control to the desired
  contact configuration

• Is there a middle ground?
• make a few assumptions about object geometry and
  configuration.
• use only one or two tactile probes to guide the
  manipulator into a goal configuration.

Dexterous Robotics Laboratory, Johnson Space
Center
3
Contact-Relative Motions

• Touch object lightly to acquire force feedback
  information with contact force sensors.
• Move manipulator relative to the contact
  positions and contact surface normals.
• Re-establish contact

Example
Requires force sensors at contacts
Dexterous Robotics Laboratory, Johnson Space
Center
4
A Sequence of Contact-Relative Motions
Dexterous Robotics Laboratory, Johnson Space
Center
5
Hardware Examples
6
Experimental Validation Grasp Quality
Dexterous Robotics Laboratory, Johnson Space
Center
7
Optimization Problem

• Given
• Action set contact-relative motions
• Set of observations derived from contact
  positions and surface normals
• Starting configuration is unknown (but
  constrained)
• A goal configuration(s)

• Calculate A policy for executing actions that
  leads to a goal configuration in the shortest
  number of steps.

• Underlying assumptions
• position control operations can be accomplished
  arbitrarily quickly.
• contact probes occur comparatively slowly.

Dexterous Robotics Laboratory, Johnson Space
Center
8
Learning contact-relative motion policies

• Observations
• Force residual
• Moment residual

• Given unknown object parameters and object pose,
  problem is partially observable
• Solve using reinforcement learning
• Resolve ambiguity by incorporating a partial
  history of state and/or actions

Dexterous Robotics Laboratory, Johnson Space
Center
9
Example grasp-the-rectangle (planar simulation)
Actions

• Oppose contact move contact 0 (resp. 1) into
  opposition with contact 1 (resp. 0)
• Implicit assumption opposition is possible

• Discrete moment residual move contact 0 (resp.
  1) to a specific moment residual configuration
  w.r.t. contact 1.
• Implicit assumption planar surface

Dexterous Robotics Laboratory, Johnson Space
Center
10
Example grasp-the-rectangle (planar simulation)
State Variables

• Moment applied by contact 0 about 1 (discretized
  into three states)
• Moment applied by contact 1 about 0 (discretized
  into three states)
• Force residual (discretized into two states)

Action history

• Which action executed?
• Did the last action succeed or fail?
• If failure, reason for failure.

Failure modes
Palm collides with object
Aperture too small
Dexterous Robotics Laboratory, Johnson Space
Center
11
Example Results (planar simulation)

• SARSA
• Learning rate 0.3
• Discount factor 0.9
• Reward -1 everywhere
• Optimistic initial values
• Random starting configurations

1.
Two strategies learned
2.
12
Hardware Implementation of Learned Policy
Dexterous Robotics Laboratory, Johnson Space
Center
13
Conclusion

• Described a class of actions that displace
  manipulator relative to contact position and
  normal.
• Demonstrated that this approach can generate good
  grasps in hardware.
• Defined optimization problem given a set of
  actions, a set of observations, and a goal
  configuration determine a policy that reaches the
  goal in the shortest number of steps.
• Solved the problem using Reinforcement Learning.
• Demonstrated practicality on hardware.

Dexterous Robotics Laboratory, Johnson Space
Center