CAPABLE: Capability-Aware Policy Adaptation via Behavioral Latent Encoding
CAPABLE adapts frozen Vision-Language-Action policies by inferring remaining actuator capability from command history and kinematics. It uses a shared temporal encoder and self-supervised physical prediction to create a latent vector that conditions a residual reinforcement learning policy. This allows the system to recover from faults without needing explicit fault labels or joint identifiers, showing strong transfer capabilities.