Megadose The AI curve, compressed.

SALA: Semantically Aligned Latent Actions Enable Dual-Axis Scaling of Embodied Foundation Models

· HF Daily Papers ·
Human video improved robot action prediction and real-world task success without paired human–robot clips.

SALA uses language descriptions to align human and robot clips by the actions they depict. This addresses a weakness in existing latent action models, whose representations can cluster by visual appearance rather than action. Robot action reconstruction anchors the shared representation in control. The authors report that more general human video reduced downstream action prediction error during pre-training, while more task-specific human video improved real-world success during post-training.

HF Daily Papers' note

score 4

Categories: Research