Safe Reinforcement Learning
Safe reinforcement learning (Safe RL) focuses on training AI agents to maximize rewards while adhering to safety constraints, addressing the inherent risk of unpredictable behavior in standard reinforcement learning. Current research emphasizes developing methods that guarantee safety through various approaches, including constraint-aware policy optimization (e.g., Lagrangian methods), shielding techniques that filter unsafe actions, and the use of learned safety models (e.g., control barrier functions, anomaly detection). These advancements are crucial for deploying RL agents in real-world applications, such as robotics and autonomous systems, where safety is paramount, and are driving progress in both theoretical understanding and practical implementation of safe and reliable AI.
Papers
Safe Reinforcement Learning Filter for Multicopter Collision-Free Tracking under disturbances
Qihan Qi, Xinsong Yang, Gang Xia
Disturbance Observer-based Control Barrier Functions with Residual Model Learning for Safe Reinforcement Learning
Dvij Kalaria, Qin Lin, John M. Dolan
Flipping-based Policy for Chance-Constrained Markov Decision Processes
Xun Shen, Shuo Jiang, Akifumi Wachi, Kaumune Hashimoto, Sebastien Gros