Safety Gym

OpenAI introduced Safety Gym, a set of simulated robot environments for studying constrained reinforcement learning and safer exploration.

The article argues that constrained RL is a better fit for safe exploration than plain reward-maximizing RL.

Safety Gym adds richer tasks and explicit cost signals, such as unsafe collisions and other safety violations.

The benchmark lets researchers test whether agents can achieve goals while respecting safety limits, not just maximizing reward.