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

Introduced REVERSAL-BENCH, a benchmark with a continuous reversibility parameter (ρ) and a reset oracle across eight manipulation settings in five physics engines. Demonstrated a sharp 'reversibility cliff' where reset-free agents fail as ρ increases, contrasting with episodic agents. Showcased that this failure mode persists in full physics simulations and is causally driven by irreversibility. Released the benchmark suite, a dataset labeled with recoverability, and a reset oracle. Evaluated a safety shield for predicting and intervening in irreversible failures.