A research team has released CRAX, a benchmarking toolkit designed to make safe reinforcement learning testing far less of a slog.
CRAX — short for Constrained RL Accelerated with JAX — runs on top of the MuJoCo XLA physics engine and uses vectorized operations with hardware acceleration to deliver up to roughly 100x speedups over comparable CPU-based safety benchmarks. The suite covers six environment types, three agent-specific tasks, and three difficulty levels each. The researchers tested six established safe RL methods against it and found that none of them consistently wins across all tasks — a result that undercuts any vendor claiming their approach "solves" safe RL.
The speed gap matters because safe RL is a prerequisite for deploying autonomous systems in the real world — robotics, self-driving vehicles, anything where a bad decision has physical consequences. Slow benchmarks create a quiet tax on researchers: fewer experiments run, fewer failure modes get caught before deployment. A 100x speedup isn't a footnote; it changes what's practical to test at all.
The study also found that curriculum learning across difficulty levels and safety transfer — training on easier settings first — outperforms throwing an agent directly at harder problems. That finding echoes what practitioners in adjacent fields have argued for years, but the field has lacked the tooling to validate it at scale. CRAX won't close every gap in safe RL research, but it removes the excuse that rigorous benchmarking is too slow to bother with.