CAPABLE: Capability-Aware Policy Adaptation via Behavioral Latent Encoding

This paper presents CAPABLE, a unified framework for adapting frozen vision-language-action (VLA) policies to physical faults. It integrates self-supervised capability inference with residual reinforcement learning, enabling bounded corrections to VLA actions without fault labels. CAPABLE improves success rates across 28 LIBERO tasks, achieving 59.3% success on actuator-excluded faults, outperforming baselines by 17.4 points.