Amap unveils whole-body foundation model, report says, using human motion data to train robots

leaderobot.com reported on Oct. 27, 2026, that Amap, an Alibaba subsidiary, had unveiled a whole-body foundation model based on open-world actions. According to the report, the model uses human action data to train robots to coordinate locomotion, posture and two-handed manipulation. Data collection for a representative task took 29 seconds, down from about 100 seconds when collected using the robot itself, representing an approximately 3.4-fold increase in collection efficiency, the report said.

The leaderobot.com report points to a difference in the data-collection process: data collection for a representative task took 29 seconds, down from about 100 seconds. The extent to which an approach using human motion can replace collection with the robot itself needs to be considered alongside the evaluation conditions and supported robot models.

The key point is not the term “whole-body foundation model” but the reported shift from collecting data with the robot itself to recording human motion for robot training. The reported reduction in collection time for a representative task, from about 100 seconds to 29 seconds, concerns the efficiency of data acquisition. That alone does not show improvements in generalization across different environments or robots, real-world task success rates, or operating costs. The scope and conditions of the evaluation, and the robot models to which the data was applied, need to be considered separately. Our Sept. 10 coverage of “Amap announces ABot-Earth 0.7, generating a 3D city in 10 minutes on a single GPU” covered 3D city-scene generation. The current report concerns linking human whole-body motion to robot training, highlighting a different layer: spatial representation versus learning bodily movements. However, the report does not establish that the two models share data or have been technically integrated. Hesai Technology’s 3D world data engine, “Kosmo” also aims to turn the real world into 3D data. How spatial and motion information are collected and fed into robot training is a relevant basis for comparing these systems. In implementation, inference at the edge, as discussed in reports on the second-generation VLA edge deployment for XPeng IRON, is a separate process from collecting training data and developing a model. The current information does not specify the computing environment in which Amap’s model runs or whether it has been installed on a particular robot. Shorter data collection therefore cannot be assumed to translate directly into edge processing or real-world tasks. Points to watch include the scale and composition of the collected data, how motions are aligned, evaluation conditions, supported robot models and performance on physical robots after training. Without those details, it is difficult to compare the approximately 3.4-fold efficiency figure across other tasks or platforms.