A new preprint on arXiv, titled Multi-View Latent Action Pretraining for Cross-Embodiment Robot Learning from Videos, argues that action-free videos can substitute for costly teleoperated data in robot learning. The authors build on latent action models (LAMs), which turn raw video into a supervision signal by inferring abstract actions between frames.
The abstract notes that existing LAMs observe a single camera, and the paper appears to extend this setup to multiple views. The motivation is that multi-view observations could provide richer supervision for learning control across different robot embodiments. However, the abstract is truncated, so specific methods and experimental results are not available in the source.
As a single-source digest, this entry only reflects the stated motivation and gap. No comparison with prior multi-view approaches or quantitative outcomes is included in the available text.