Researchers have published HumanVerse, a 500‑hour collection of egocentric video paired with full‑body and hand motion. They also released λ0, a humanoid foundation model trained on this data. λ0 was tested on the Unitree G1 robot and reached 90.0% success on the SIMPLE benchmark and 62.5% on real‑world loco‑manipulation tasks, all without robot‑specific demonstrations.
Why it matters
The dataset and model let robot builders teach whole‑body manipulation to humanoid robots without collecting costly robot demonstrations.
