kapynResearch

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

REVERSAL-BENCH is a benchmark that measures how well reinforcement learning agents can learn without external resets. It introduces a continuous reversibility parameter ρ and a reset oracle to verify state recoverability across eight manipulation tasks in five physics engines. The framework gives developers a systematic way to evaluate agents’ robustness in non‑reversible real‑world scenarios, a key hurdle for autonomous systems.

Apple ML Research·Sep 17, 2026

Opening Kapyn…