PAC-NMPC with Learned Perception-Informed Value Function

Nonlinear model predictive control (NMPC) is typically restricted to short, finite horizons to limit the computational burden of online optimization. This makes a global planner necessary to avoid local minima when using NMPC for navigation in complex environments. For this reason, the performance of NMPC approaches are often limited by that of the global planner. While control policies trained with reinforcement learning (RL) can theoretically learn to avoid such local minima, they are usually unable to guarantee enforcement of general state constraints. In this paper, we augment a sampling-based stochastic NMPC (SNMPC) approach with an RL trained perception-informed value function. This allows the system to avoid observable local minima in the environment by reasoning about perception information beyond the finite planning horizon. By using Probably Approximately Correct NMPC (PAC-NMPC) as our base controller, we are also able to generate statistical guarantees of performance and safety. We demonstrate our approach in simulation and on hardware using a 1/10th scale rally car with lidar.