Pandaily reported on 24 October that Amap, Alibaba Group’s mapping and location unit, and Alibaba’s ATH business group had released Lambda-0, a whole-body humanoid foundation model trained on human first-person data rather than robot teleoperation recordings.
The team collected 500 hours of video, body movement and hand-action data from people wearing consumer-grade devices in everyday spaces, with no robot present. It trained the model in three stages, moving from hand interactions to whole-body coordination and then to controlling a humanoid. Viewpoint matching and a scheme that adjusts actions from broad movements to finer ones address the differences between human and robot bodies, including errors in where a robot’s hands make contact.
Moving things across a room would involve walking and handling objects together. The approach could allow those movements to be learned from human recordings without collecting demonstrations on a physical robot.
On seven whole-body and longer mobile-manipulation tasks, the team reports a 62.5 per cent success rate when transferring Lambda-0 to a Unitree G1 without additional robot-specific training, 22.5 percentage points above its strongest baseline. It also reports about 90 per cent on the SIMPLE simulation benchmark. The code is available on GitHub under an Apache 2.0 licence, alongside an arXiv paper. The team’s results have not yet been independently reproduced, Pandaily reported.
