Learning to Solve Long-Horizon Robot Manipulation Problems

Abstract

If we want mobile robots that perform multi-step tasks in visually diverse and geometrically complex environments, we need them to quickly decide what to do and how to do it. Manipulating multiple objects in environments with movable and articulated obstacles over time requires the robot to satisfy constraints like collision-freeness, reachability, and action feasibility. For problems with large state spaces, continuous action spaces, and long decision horizons, the hybrid constraint satisfaction problems induced by planners become combinatorially difficult to solve. In this thesis, I will discuss strategies for using offline learning to speed up deploymenttime planning, i.e., using a plan feasibility predictor, a subgoal generator, or a compositional joint continuous constraint solver. I will also present strategies for chaining policies learned from demonstrations using conditional inputs, such as key poses and natural language, for generalization in real-world environments. With the resulting efficient long-horizon manipulation planning system, we can solve complex robotic manipulation problems faster at deployment time. It can also be used to generate diverse large-scale whole-body trajectories as part of the data mixture for training robot foundation models in embodied reasoning, planning, and acting.