Action-Similarity Supervision Improves Cross-Embodiment Transfer in Latent Action Models
September 18, 2026
Latent action models (LAMs) often fail to encode identical motions from different robots into consistent latent spaces. This method uses action-similarity supervision to force the similarity between two latent actions to match the similarity of their corresponding ground-truth robot action sequences without directly predicting the actions.
HOW THIS AFFECTS YOU
●
builderThis offers a path to training more generalist robot policies using video data from diverse embodiments.
●
researcherYou can improve latent space consistency across different robot hardware using similarity-based loss.