Safe Learning via Warm-starting

When we learn new information about the world, we need to update our safety analysis to maintain guarantees. Unfortunately, the compute required to recompute these analyses can be intractable for realistic applications. We work on methods to quickly adapt our knowledge of safety based on changes in the environment, other agents, model-mismatch, and external disturbances. These methods maintain convergence (and therefore safety) guarantees.

Relevant Papers:

S Tonkens, SL Herbert. Refining Control Barrier Functions through Hamilton-Jacobi Reachability. IEEE Conference on Intelligent Robots and Systems, 2022.

SL Herbert*, Jason Choi*, Suvansh Sanjeev, Marsalis Gibson, Koushil Sreenath, CJ Tomlin. Scalable Learning of Safety Guarantees for Autonomous Systems using Hamilton-Jacobi Reachability. IEEE International Conference on Robotics and Automation, 2021.

SL Herbert, S Bansal, S Ghosh, CJ Tomlin. Reachability-based safety guarantees using efficient initializations. IEEE 58th Conference on Decision and Control (CDC), 2019.