Human-collected robot manipulation data are rarely clean. They mix successful actions with retries, pauses, recoveries, and outright failures. This poses a problem for standard behavior cloning, which typically assumes demonstrations are consistent and goal-directed.

Advantage-Conditioned Behavior Cloning (AC-BC) is a promising alternative because it can exploit mixed-quality data by conditioning the policy on an advantage estimate. However, the arXiv preprint argues that AC-BC only works if the way this condition is estimated aligns with how it is expressed during policy learning. The proposed method, ALIGN, is designed to correct that misalignment.

The paper is presented as a conceptual and methodological contribution, with the abstract focusing on the problem and the proposed solution. Since the source is a single preprint and the abstract is truncated, no experimental results or comparisons are available here; the significance lies in identifying and addressing a key failure mode in advantage-conditioned cloning.