Estimating Control Barriers from Offline Data

Abstract

Learning-based control methods must satisfy safety requirements in order to be deployed in real-world robotics systems. Control barriers, a potential candidate for standardizing the notion of safety in the learning community, achieve theoretical guarantee of controller safety via specifying forward-invariant safe region of the system using Lyapunov theory. In this work, we develop a model-based learning approach to synthesize robust safety-critical controllers by constructing neural control barriers solely on offline data. An actor model is learned to capture the safest controls with respect to neural control barriers. Then we incorporate the actor model in optimizing the derivative of the barrier model to satisfy Lyapunov condition so that no optimality assumption is imposed to the controls from data. The actor also enables us to annotate unlabeled data, e.g. the demonstrations with questionable safety, via out-of-distribution analysis. This is to best serve the offline setting that targets to learn directly from real-world demonstrations with limited labeled data. We evaluate the proposed algorithm for obstacle avoidance in both simulation and real-world platform. Trained on a limited amount of real-world data, the new method can achieve comparable performance to the DWB local planner included with ROS2 Nav2 for static obstacle avoidance, as well as handle dynamic obstacle avoidance from sensory data on real-world platform.

Type
Publication
IEEE International Conference on Robotics and Automation