Robust Closed-Loop Manipulation via Contact Sensing
Contact sensors can provide robots with with the feedback vital to robustly manipulating objects in uncertain environments. However, there are three principal challenges to using contact sensing. First, contact is inherently discontinuous. Second, contact sensors only provide rich information while in contact. Third, planning for contact requires reasoning about the physics of interaction.
Our work addresses these challenges by formulating contact manipulation as a partially observable Markov decision process (POMDP) in the joint space of robot configuration and object poses. Policies generated from this POMDP naturally take information-gathering actions when necessary to complete the task. For example, the robot may force an object into a contact sensor to localize it before attempting to push it into the goal region.
We specifically consider the case of quasistatic planar manipulation. We exploit the structure of the problem to quickly find a near-optimal policy:
- Pre- and Post-Contact Policy Decomposition. We show that the optimal policy naturally decouples into pre- and post-contact stages. We present an algorithm that leverages this insight to reuse one post-contact policy across multiple problem instances.
- Configuration Lattice. We perform an online search in a lazily-constructed configuration space lattice to guarantee that all actions the robot takes are feasible; e.g. do not exceed joint limits, violate kinematic reachability constraints, or collide with obstacles.
- Heuristics. We guide an online search in full state space with heuristics derived from relaxations of the problem. For example, the optimal policy for an empty environment is a useful heuristic to follow when manipulating objects in moderate amounts of clutter.
We validate the efficacy of our approach in simulation and real-robot experiments on HERB, a manipulator with a 7-DOF Barrett WAM arm and a BarrettHand end-effector. We show that our approach successfully completes a planar pushing task more often than baseline algorithms that do not take information-gathering actions. Finally, through my collaboration with NASA, I demonstrates that this approach generalizes between robots by using the same algorithm to generate successful policies for Robonaut 2, an anthropomorphic humanoid robot.
Publications
- Michael C. Koval. Robust Manipulation via Contact Sensing. PhD thesis, Carnegie Mellon University, 2016. PDF
- Michael C. Koval, David Hsu, Nancy S. Pollard, and Siddhartha S. Srinivasa. Configuration lattices for planar contact manipulation under uncertainty. In Workshop on the Algorithmic Foundations of Robotics. December 2016. PDF.
- Michael Koval, David Hsu, Nancy Pollard, and Siddhartha Srinivasa. Configuration lattices for planar contact manipulation under uncertainty. arXiv:1605.00169, May 2016. Preprint. PDF
- Michael Koval, Nancy Pollard, and Siddhartha Srinivasa. Pre- and post-contact policy decomposition for planar contact manipulation under uncertainty. International Journal of Robotics Research, 35(1-3):244–264, 2016. PDF
- Michael Koval, Nancy Pollard, and Siddhartha Srinivasa. Pre- and post-contact policy decomposition for planar contact manipulation under uncertainty. In Robotics: Science and Systems. July 2014. PDF
Videos