Safe RL

Safe reinforcement learning aims to maximize task performance while limiting hazards such as collisions, equipment damage, or unsafe medical decisions, including during exploration and deployment. Common approaches formulate constraints as costs in constrained Markov decision processes and use primal-dual or Lagrangian optimization; others add control-barrier-function, Lyapunov, or reachability shields that filter actions, sometimes using learned models to account for uncertainty. Work also addresses offline learning, human preference feedback, temporal-logic specifications, and applications from robotics and driving to healthcare and energy control. A central challenge is balancing strong, reliable safety guarantees with reward, data efficiency, and adaptation to unfamiliar conditions; benchmarks also show that methods vary across tasks and that average costs can hide rare severe failures.