Error Amplification in Action Chunking Policies

A study of action chunking in behavioral cloning reveals that error amplification is common during execution. By injecting perturbations, the research finds that confidently stable states are rare across twelve manipulation tasks, challenging the assumption of temporal consistency in open-loop regimes.