Risk-Gated RL for Safety-Critical Control Under Partial Observability

A novel reinforcement learning technique known as action-conditioned risk gating tackles the challenge of risk-sensitive control within partially observable Markov decision processes. This method employs a concise finite-history proxy state and develops a predictor for imminent safety violations. The anticipated risk functions as a penalty during the value learning phase and acts as a decision-time gate, balancing between optimistic and conservative ensemble value assessments, which allows for the efficient evaluation of low-risk actions. The design aims to lower both computational expenses and model sensitivity in comparison to belief-space planning.