Aligning robot actions with condition signals improves manipulation success
ALIGN: Aligning Condition Estimation and Expression in Advantage-Conditioned Behavior Cloning
RoboticsArtificial Intelligence
Summary
Robot data often mixes good attempts with retries and failures, making it hard for robots to learn from it. The authors created ALIGN, a method that keeps condition signals clear and directly connected to robot actions. This helps robots better understand when to try harder, pause, or recover. Their tests with real robots stacking bowls and hanging keys showed much higher success rates than previous methods. ALIGN helps robots learn reliable skills from messy real-world data.
What this means in practice
- •For robotics engineers: Extract robust manipulation policies from mixed-quality real-world robot data by better aligning condition signals with action predictions.
- •For industrial automation teams: Improve success rates in assembly or handling tasks by integrating retry and recovery data effectively during robot policy training.
Authors
Zechen Wang, Zhenqi Qiu, Houlong Xiong, Suohang Zhang, Yiyu Ren, Wanting Xu, Hongfei Niu, Chengyang He, Ge Sun, Ran Cheng, Qian Zhu
Abstract
Human-collected robot manipulation data often mix successful behavior with retries, pauses, recoveries, and failures. 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. Existing pipelines can weaken this signal by thresholding estimator-derived advantage scores into noisy conditions, and by routing conditions through prompts that weakly affect the action vector field. We introduce ALIGN, a condition-aligned AC-BC framework for vision-language-action policies. ALIGN directly predicts categorical pairwise conditions in the policy-facing label space and injects them into a DiT action expert through AdaRMS modulation fused with flow time. 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. ALIGN improves condition fidelity, strengthens conditional vector-field separability, and achieves 90.0% and 56.7% closed-loop success, compared with 56.7% and 33.3% for the strongest AC-BC baseline. 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.