Safe exploration in Reinforcement Learning
Matteo Turchetta
Deploying autonomous agents in safety-critical, real-world applications requires the ability to account for safety constraints, since the environment or the dynamics are often unknown a priori. We present some of the recent work we have been doing at the Learning and Adaptive Systems at ETH in order to address the problem of safe exploration. In particular, we introduce different notions of safety that are suitable for different problems. Moreover, we present a set of algorithms leveraging ideas from graph search methods, Lyapunov stability theory and model predictive control that can provably achieve safe exploration.