Learning potential functions by demonstration for path planning

Potential functions can be used to design efficient path planning schemes. However, it is often difficult to design appropriate potential functions to mimic desired behavior of the agent. Instead of using a pre-designed potential function for path planning, this paper presents an algorithm that learns the underlying potential function from a given sample trajectory generated by a “expert” (say, a human). This underlying potential function implicitly incorporates obstacle avoidance information that may be intuitive or experience-based. The potential function to be learned is parametrized and the parameter weights are obtained through minimization of a well-designed cost function via a gradient descent search algorithm. Once learned, this potential function can be used for path planning in case of alternative (and more complex) scenarios, such as those with multiple obstacles. The paper presents the theoretical foundation and numerical validation of the proposed algorithm.