Generative Planning with Fast Collision Checks for High Speed Navigation

Reasoning about large numbers of diverse plans to achieve high speed navigation in cluttered environments remains a challenge for robotic systems even in the case of perfect perceptual information. Often, this is tackled by methods that iteratively optimize around a prior seeded trajectory and consequently restrict to local optima. We present a novel planning method using normalizing flows (NFs) to encode expert-styled motion primitives. We also present an accelerated collision checking framework that enables rejecting samples from the prior distribution before running them through the NF model for rapid sampling of collision-free trajectories. The choice of an NF as the generator permits a flexible way to encode diverse multi-modal behavior distributions while maintaining a smooth relation to the input space which allows approximating collision checks on NF inputs rather than outputs. We show comparable performance to model predictive path integral control in random cluttered environments and improved exit rates in a cul-de-sac environment. We conclude by discussing our plans for future work to improve both safety and performance of our controller.