Configuration Lattices for Planar Contact Manipulation Under Uncertainty

Abstract

This work addresses the challenge of a robot using real-time feedback from contact sensors to reliably manipulate a movable object on a cluttered tabletop. We formulate this task as a partially observable Markov decision process (POMDP) in the joint space of robot configurations and object poses. This formulation enables the robot to explicitly reason about uncertainty and all major types of kinematic constraints: reachability, joint limits, and collision. We solve the POMDP using DESPOT, a state-of-the-art online POMDP solver, by leveraging two key ideas for computational efficiency. First, we lazily construct a discrete lattice in the robot’s configuration space. Second, we guide the search with heuristics derived from an unconstrained relaxation of the problem. We empirically show that our approach outperforms several baselines on a simulated seven degree-of-freedom manipulator.