Gemini Robotics 2 claims whole-body intelligence
New physical-AI models aimed at coordinating a whole robot rather than a manipulator. Whole-body control is the specific thing humanoids have been bad at, so the claim is well aimed — and hard to verify from outside.
Google DeepMind has launched Gemini Robotics 2, described as bringing whole-body intelligence to robots.
The phrase names a real gap. Most robot learning has been about manipulation — a hand, an arm, a grasp — because manipulation is where the datasets and the benchmarks are. Whole-body control means balance, gait, contact with the environment and the coordination between all of it, and it is the part where humanoids visibly struggle in the field.
It is also the part that is hardest to evaluate from outside. A manipulation result can be reported as a success rate on a task. Whole-body competence shows up as not falling over on a surface nobody photographed, and there is no leaderboard for that.
What would make the claim checkable is deployment duration rather than demonstration. The units that have accumulated real field time are the ones whose failure modes are known, and those failure modes are exactly what whole-body control is supposed to address.
Elsewhere in the same week, Generative Bionics launched Gene.01, a smart-skin humanoid platform for safe human collaboration, released with an open-source robot model. Contact sensing across the whole surface is the other half of this problem, approached from hardware instead of policy.
Humanoid Press — Physical AI & Embodied AI Daily News → · HumanoidHub — Humanoid Robotics News & Updates →