Alibaba's Amap unit launches robot foundation model for human-like movement
Alibaba's mapping division, Amap, has developed a new model dubbed λ0, designed to train humanoid robots to execute complex tasks by analyzing human movement from video data.
The new foundation model, λ0, aims to bridge the gap between human physical activity and robotic execution. Developed by Alibaba’s Amap and Token Hub, the system utilizes first-person video coupled with motion data to teach humanoid robots how to synchronize walking, posture adjustments, and manual manipulation.
To facilitate this training, the research team created HumanVerse, a dataset comprising 500 hours of real-world human activity. By analyzing these recordings, the model enables robots to learn complex maneuvers without the need for traditional, labor-intensive teleoperation methods. In initial testing, the model recorded a 62.5% success rate in combined locomotion and manipulation tasks.
This move aligns with broader efforts in the robotics industry to increase the autonomy of humanoid systems in unstructured environments. Alibaba has opted to release sample data and hardware configuration code, providing a baseline for external researchers to experiment with the λ0 framework.