Structured learning and inference for robot motion generation

Abstract

The ability to generate motions that accomplish desired tasks is fundamental to any robotic system. Robots must be able to generate such motions in a safe and feasible manner, sufficiently quickly, and in dynamic and uncertain environments. In addressing these problems, there exists a dichotomy between traditional methods and modern learning-based approaches. Often both of these paradigms exhibit complementary strengths and weaknesses, for example, while the former are interpretable and integrate prior knowledge, the latter are data-driven and flexible to design. In this thesis, I present two techniques for robot motion generation that exploit structure to bridge this gap and leverage the best of both worlds to efficiently find solutions in real-time. The first technique is a planning as inference framework that encodes structure through probabilistic graphical models, and the second technique is a reactive policy synthesis framework that encodes structure through task-map trees. Within both frameworks, I present two strategies that use said structure as a canvas to incorporate learning in a manner that is generalizable and interpretable while maintaining constraints like safety even during learning.