Google DeepMind is training Gemini Robotics 2 to control several kinds of robot body, including Apptronik’s Apollo 2 humanoid. The ambitious bit is portability: skills learned in one setting could eventually carry across machines, although today’s test results still look more apprentice than Rosie the Robot.
Google DeepMind Watch analysis
What happened
Shown in late July, Gemini Robotics 2 extends DeepMind’s earlier work from mainly upper-body control to coordinated movement involving legs, torso, arms and fingers. Demonstrations include fetching snacks, changing lightbulbs and tying knots across humanoids and smaller research platforms.
The software uses two layers. Gemini Robotics ER 2 turns a spoken request into steps and monitors progress, while a lower-level vision-language-action model translates instructions and camera input into movements, updating its predictions four or five times each second.
Key findings
- One model family, several bodies
DeepMind wants learned skills to transfer across robot shapes instead of requiring a separate intelligence for every machine. - Whole-body control is now in scope
Gemini Robotics 2 coordinates legs, torso, arms and fingers rather than concentrating largely on movement above the waist. - Planning and movement are separated
ER 2 breaks down requests and tracks completion while a lower-level model handles the robot’s immediate physical actions. - Progress tracking remains imperfect
DeepMind reported 57.4% accuracy when ER 2 classified how far a task had progressed. - Ordinary chores remain stubbornly difficult
Apptronik’s Apollo completed DeepMind’s dustpan test only 32% of the time. - Touch is a substantial blind spot
Internet-scale visual data cannot supply the tactile, force and proprioceptive feedback humans use when handling grapes, bags or glass.
Why it matters
A robot intelligence that transfers between bodies could reduce the need to engineer every skill separately for every machine. That would make robots more adaptable in homes, factories and other untidy environments where rigidly programmed motions quickly meet reality and lose the argument.
The limitations are equally instructive. Cameras can tell a model that it is holding a grape, but not reliably how close the grape is to bursting. Scientific American’s account shows that general-purpose robotics will require better physical feedback, more varied training data and much more consistent execution, not merely a larger language model bolted to a pair of legs.
Our read
This is a serious attempt to give robots transferable intelligence rather than ever-longer lists of choreographed tricks. The architecture makes sense, the demonstrations are promising and the whole-body expansion is meaningful.
It is not yet a general-purpose robot brain. A 32% result on sweeping debris is a splendid antidote to any imminent-robot-butler fantasies. Developers and buyers should judge these systems by repeatable task completion, recovery from mistakes and safe behaviour around people, not by the most photogenic successful run.
What to watch
- Whether skills transfer reliably between substantially different robot bodies.
- Improvements in tactile sensing and training data for fragile or deformable objects.
- Independent reproduction of DeepMind’s task-success and safety results.
- Whether Apollo 2 progresses from demonstrations to sustained real-world deployments.
Discussion spark: Which capability would most convincingly show that a robot foundation model can generalise beyond a polished demonstration?
Sources and evidence
- Google DeepMind wants to build an AI brain that can jump between robot bodies – Scientific American (15 September 2026, 10:00 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind