Event

Proposes embodiment-classifier guidance for demonstration-free transfer of a frozen VLA to new embodiments

1 reports · 1 sources

Devin Han and Hershraj Niranjani propose the Embodiment-Classifier Guidance method: train a lightweight embodiment classifier to distinguish successful action chunks on the source embodiment from those on the target embodiment, then use the classifier's gradients to guide a frozen π0 policy during denoising steps, enabling it to transfer to a new robot embodiment without demonstrations or parameter updates. On the LIBERO simulation benchmark with four modified controllers, this guidance raises π0's average success rate on the target embodiment from 72% to 80% (12 suite-controller pairs); retraining the classifier with successful samples obtained via guidance raises the success rate from 52.5% to 70.7%. However, guidance performs inconsistently on weak wrist joints and is harmful when the wrist joint is locked—this occurs when the classifier lacks data or prefers infeasible actions. The authors state that the method is data-efficient in simulation, requiring only a small number of target-embodiment rollouts per task, and therefore has potential for sim-to-real transfer where rollouts are expensive.