A Bayesian Active Learning Approach to Adaptive Motion Planning

摘要

Motion planning, the task of computing collision-free motions for a robotic system from a start to a goal configuration, has a rich and varied history. Up until now, the bulk of the prominent research has focused on the development of tractable planning algorithms with provable worst-case performance guarantees such as computational complexity, probabilistic completeness or asymptotic optimality. In contrast, analysis of the expected performance of these algorithms on the real world planning problems a robot encounters has received considerably less attention, primarily due to the lack of standardized datasets or robotic platforms. However, recent advances in affordable sensors and actuators have enabled mass deployment of robots that navigate, interact and collect real data. This motivates us to examine new algorithmic questions such as: "How can we design planning algorithms that, subject to on-board computation constraints, maximize their expected performance on the actual distribution of problems that a robot encounters?"