Oliver Hausdörfer and his team released a new latent action model that uses synchronized views from two cameras. The model infers 6‑degree‑of‑freedom actions more accurately, raising decoding accuracy by 17% over single‑view versions. When plugged into a robot policy, success rates on benchmark tasks rose 13.5%, compared with 6.2% for the single‑view baseline. In real‑world tests with a robot copying human‑hand demonstrations, the dual‑view model added 20% more successful rollouts across 850 attempts.
Why it matters
The improvement means robots can learn new skills from ordinary videos with fewer demonstrations, helping labs with limited tele‑operation data.
