Gemini Robotics 2: Google Now Controls a Humanoid From Feet to Fingertips

Earlier Gemini Robotics demonstrations concentrated heavily on upper-body and tabletop tasks. Gemini Robotics 2 expands the action space to the entire humanoid. A command can now require the robot to walk toward a shelf, change its center of mass, bend or crouch, position an arm and manipulate the target while maintaining balance.

That coupling is technically important. Locomotion and manipulation cannot always be solved as independent modules because reaching changes the body's stability and the body pose changes which grasps are possible. A useful whole-body policy has to reason about the task while respecting those coupled constraints.

DeepMind's example asks Apptronik Apollo 2 to move a watering can to a green bin on a bottom shelf. The robot walks to the object, grasps it, takes several steps and places it at the destination. The point is not the watering can; it is the transition from language to a sequence that spans navigation and manipulation.

DeepMind also notes that movement speed still needs improvement. That caveat matters commercially. A task that succeeds at low speed in a lab may not meet production cycle time, and faster whole-body motion narrows the control margin around balance and contact.

With Apollo 2 using Inspire hands, DeepMind reports 68.4% for picking from a table, 45.7% from the floor and 76.3% from a shelf. The floor result is a useful reminder that a seemingly small geometric change can force a different stance, reach envelope and balance strategy.

On Franka Duo, two-finger grippers perform better on several precision tasks: DeepMind reports 74.2% for general pick-and-place, 78.9% for diverse tool kitting and 89.6% for precise insertion. These are developer-reported averages under DeepMind's task protocols, not a universal robot ranking.

Gemini Robotics 2 controls a five-finger SharpaWave hand with 22 degrees of freedom on Apollo 2. DeepMind shows knot tying and sealing a ziplock bag. Each finger adds possible contacts, and contact changes what the next joint motion should be. The controller has to handle underactuated objects, occlusion and several acceptable grasp configurations.

DeepMind explicitly states that multifinger dexterous manipulation remains challenging. That distinction is essential. A model can understand the language instruction and still fail because fingertip pose, friction, force or timing is wrong by a small amount.

DeepMind shows the same model checkpoint controlling Apollo 2 with different end effectors and a Franka Duo. That is meaningful evidence that a shared model can span distinct action interfaces when the training and system design account for them.

It should not be read as plug-and-play control of any unseen robot. Kinematics, joint limits, cameras, control rates and safety interfaces still differ. DeepMind's On-Device 2 work separately addresses adaptation to new bi-arm embodiments, which reinforces the point that embodiment transfer usually needs an explicit adaptation mechanism.

The next evidence should include longer runs, disturbance recovery, intervention rates, control latency, repeated trials on unseen layouts and performance under human proximity. Whole-body capability increases the value of the robot but also increases the space of unsafe or unstable motions.

Gemini Robotics 2 is best read as a major expansion of the learned action space. The remaining problem is reliability: whether these abilities can run fast, repeatedly and safely enough to beat a more specialized automation stack on real cost and uptime.