REVERSAL-BENCH: A Reversibility Axis and Reset Oracle for Measuring the Reset-Free RL Cliff

Researchers introduced REVERSAL-BENCH, a benchmark to test reinforcement learning agents' ability to recover from irreversible states in manipulation tasks without resets. They found a sharp failure point where agents get trapped in irrecoverable states as environmental reversibility decreases, halting learning. The benchmark includes multiple physics engines, a reset oracle, and a dataset to evaluate and improve reset-free RL policies.