The foundation model learns to walk, adjust its posture and manipulate with two hands from videos to the first person, and reaches 62.5% success in real-life conditions.
Amap, Alibaba's mapping service, and Alibaba Token Hub unveiled λ0 (lambda zero), a foundation model for humanoid robots that learns coordinated movements of the whole body by observing humans. It combines locomotion, posture adjustments and bimanual manipulation in the same model.
His training is based on HumanVerse, a data set of 500 hours of real human activities, filmed at the first person and synchronized with captures of body and hand movements. Interest: reducing reliance on robot teleoperation, a costly and slow collection method that remains the main bottleneck in robotic learning. On a set of real tasks combining displacement and manipulation, λ0 has a success rate of 62.5%.
The team makes available to researchers the physical configuration and code of its collection device, as well as data samples.
Read the full article on the original source
Go to the source →