Gemini Robotics 2 controls legs, torso, arms and hands under one policy
DeepMind extended Gemini Robotics on 30 July with its first model to control a humanoid's legs, torso, arms and hands under a single policy. Whole-body control from one model, rather than a locomotion stack bolted to a manipulation stack, is a different engineering proposition.
The split architecture existed for a good reason. Locomotion is a fast, high-frequency balance problem; manipulation is a slower, precision problem; and separating them let each be solved with appropriate methods. The cost was that behaviours requiring both at once — bracing against a surface, using body weight to apply force, catching balance while carrying something — fell into the gap between two controllers.
A single policy closes that gap in principle and opens a data problem in practice. Whole-body control means the training distribution has to cover coordinated whole-body behaviour, and that is precisely the data the field has least of. Teleoperation collects manipulation well and coordinated dynamic movement poorly.
Which is why the same constraint shows up in every model in this category. Gemini Robotics, NVIDIA GR00T, Physical Intelligence, Figure Helix and Skild Brain differ in architecture and agree entirely on the bottleneck: too little robot data, with answers split between mass teleoperation, shared datasets like Open X-Embodiment, and synthetic data from world models.
TechTimes — Gemini Robotics 2 controls full humanoids: legs, torso, arms and fingers under one policy → · GCN — Google DeepMind launches Gemini Robotics 2 with full humanoid body control → · TUN — Google DeepMind's Gemini Robotics 2 teaches robots full-body control →