Learning Adaptive Sampling Distributions for Motion Planning by Self-Imitation

Sampling based motion planning algorithms are widely used due to their effectiveness on problems with large state spaces by incremental tree growth in conjunction with uniform, random sampling. The major bottleneck in the performance of such algorithms is the amount of collision checks performed, which in turns depends on the sampling distribution itself. In this work, we present a framework to learn an adaptive, non-stationary sampling distribution which explicitly minimizes the search effort, given by the amount of collision checks performed. Our framework models the sequential nature of the problem by leveraging both the instantaneous search tree over the robot configuration space, as well as the workspace environment, by encoding them with a conditional variational auto-encoder, to learn a stochastic sampling policy. We encode the workspace environment with a convolutional network, and the configuration space planning tree with a recurrent neural network. We introduce an approximate oracle which can return multiple label samples for a partially solved planning problem, by forward simulating it. We use an imitation via iterative supervised learning framework to learn a stochastic sampling policy. We call this self-supervised imitation of an oracle generated by forward simulation as self-imitation. We validate our approach on a 4D kinodynamic helicopter planning problem with glideslope and curvature constraints, and a 2D holonomic problem.