CRAX: Fast Safe Reinforcement Learning Benchmarking
Researchers have developed CRAX, a new benchmark for reinforcement learning (RL) agents designed to accelerate safety testing in real-world applications. Built on the MuJoCo XLA physics engine, CRAX offers up to a 100x speedup compared to existing benchmarks, enabling more extensive experimentation. The benchmark includes six environment suites and three agent-specific tasks with varying difficulty levels. Initial evaluations of six popular safe RL methods revealed that no single method consistently outperformed others, highlighting the trade-offs between performance and safety, and suggesting that curriculum learning and safety transfer can enhance results in more challenging scenarios. AI
IMPACT Enables faster and more extensive safety testing for real-world RL applications, potentially accelerating deployment in robotics and autonomous driving.