The authors collected open-world human demonstrations using egocentric cameras and lightweight wearable motion sensors.
Human demonstrations scale whole-body learning for humanoid loco-manipulation
The authors combine 500 hours of synchronized egocentric human activity with embodiment alignment and robot training to improve Unitree G1 control.
Chinese Tech
Lambda-humanoid Team
Alibaba Token Hub · Amap · Alibaba Group
Research Digest··3 min read
The Lambda-humanoid Team presents HumanVerse, a 500-hour dataset pairing first-person video with synchronized body and reconstructed hand motion, plus λ₀, a vision-language-action policy trained to convert that experience into humanoid behavior.
Why this paper
From Alibaba Token Hub and 2 others
In one line
Robot-aligned whole-body human demonstrations scale humanoid loco-manipulation training, improving task progress and producing strong benchmark and real-world results.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (3 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§