ALIGN: Aligning Condition Estimation and Expression in Advantage-Conditioned Behavior Cloning

Key points

• Human-collected robot manipulation data often mix successful behavior with retries, pauses, recoveries, and failures.1
• Advantage-Conditioned Behavior Cloning (AC-BC) can exploit such mixed-quality data, but only if the conditioning signal is both accurately constructed from offline trajectories and effectively incorporated into the policy's action-vector-field prediction.1
• We evaluate ALIGN on deliberately mixed real-robot datasets for ARX X5 bowl stacking and Franka key hanging, each combining standard demonstrations with retry-heavy trajectories.1
• These results indicate that aligning conditions from estimation to action prediction is an effective way to extract robust manipulation policies from data in which success, recovery and error coexist.1

Sources (1)

• [1]ALIGN: Aligning Condition Estimation and Expression in Advantage-Conditioned Behavior Cloning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 02:10 PM

  We introduce ALIGN, a condition-aligned AC-BC framework for vision-language-action policies.

  Human-collected robot manipulation data often mix successful behavior with retries, pauses, recoveries, and failures.

Extractive summary: sentences quoted from the sources.