Long-Horizon Planning Under Uncertainty and Geometric Constraints for Mobile Manipulation by Autonomous Humanoid Robots

Abstract

Autonomous humanoid robots have the potential to perform critical and labor-intensive tasks that could go a long way to improve upon the quality of human life. To realize this potential, an autonomous humanoid robot must be capable of planning the right set of long-horizon actions under the conditions of uncertainty and geometric constraints that characterize real-world environments.

This thesis proposes long-horizon planning approaches for humanoid robots under conditions of uncertainty and geometric constraints that are typical of real-world environments. The specific contributions of this thesis are, 1) A reactive and efficient task planning approach for planning under low-entropy conditions in the robot's belief of the state of the world, 2) A reactive and probabilistic long-horizon planning approach for long-horizon tasks under state estimation and action uncertainty and 3) An optimal long-horizon planning approach for geometrically constrained tasks based on mixed integer convex programming.

We demonstrate the effectiveness of the approaches presented in this thesis on object rearrangement and mobile manipulation tasks in a domestic environment using the Agility Robotics Digit Bipedal Humanoid Robot and evaluate the presented approaches on planning time and task success rate metrics.