Safe Planning via Model Predictive Shielding

摘要

Reinforcement learning is a promising approach to synthesizing policies for robotics tasks. A key challenge is ensuring safety of the learned policy---e.g., that a walking robot does not fall over, or an autonomous car does not run into an obstacle. We focus on the setting where the dynamics are known, and the goal is to prove that a policy trained in simulation satisfies a given safety constraint. We build on an approach called shielding, which uses a backup policy to override the learned policy as needed to ensure safety. Our algorithm, called model predictive shielding (MPS), computes whether it is safe to use the learned policy on-the-fly instead of ahead-of-time. By doing so, our approach is computationally efficient, and can furthermore be used to ensure safety even in novel environments. Finally, we empirically demonstrate the benefits of our approach.