Safe Learning in the Real World via Adaptive Shielding with Hamilton-Jacobi Reachability

We present a robust shielding framework using Hamilton-Jacobi Reachability that can be combined with any off-policy Reinforcement Learning to enable safer learning. Using an approximate model of a system dynamics, our method can capture the local model mismatch from a safety perspective. This leads to a more conservative safety filter that can adapt to model mismatch. Using a Turtlebot 2, we demonstrate that our method can allow for safe learning in the real-world with minimal human intervention.