Amap Releases Whole-Body Foundation Model for Humanoid Robot Learning

Amap, part of Alibaba, has released what it describes as the first whole-body foundation model built on open-world behaviors. The model relies only on human action data, allowing humanoid robots to learn coordinated movement, posture and two-handed manipulation as an integrated whole.

In multiple simulation and real-robot evaluations, the model delivered leading performance, according to the company. Its data collection efficiency is about 3.4 times that of conventional methods, suggesting a lower-cost path for gathering the large-scale motion data needed to train embodied systems.

The release points to a broader shift in embodied AI: rather than relying solely on robot-collected teleoperation data, researchers can leverage human activity to build a foundation for whole-body control. Amap's model is positioned as a step toward a more scalable data base for humanoid robots; the company says the approach could reduce reliance on costly robot teleoperation and expand training data. Independent validation and broader real-world testing remain important.