Gemini Robotics 2 lands; ER 2 reasoning model live in AI Studio
Three models bring whole-body control and multi-robot teamwork to humanoids. The reasoning model is in AI Studio today; the rest stay partner-gated.
Copy markdown
Three models, one whole-body brain
Google DeepMind shipped Gemini Robotics 2 (the vision-language-action model that drives full humanoids and bi-arm robots), Gemini Robotics-ER 2 (embodied reasoning and planning), and Gemini Robotics On-Device 2 (local, adapts to a new robot in hours). The headline capability is whole-body control: a humanoid can walk to a table, pick up a watering can, cross the room, and place it precisely.
ER 2 is the part you can use today
The embodied-reasoning model, Gemini Robotics-ER 2, is available now in Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform. That is the real developer door: spatial reasoning, task planning, and human dialogue you can call today, no robot required to prototype the 'brain.' The VLA and On-Device models still need early-access partner signup.
Dexterity, with the numbers Google didn't hide
Whole-body manipulation lands 45.7%-76.3% success across tasks; standard grippers hit 74.2%-89.6%; the 22-DOF multi-finger hands swing 32%-92%, with knot-tying and ziplock-sealing still hard. These are honest per-task success rates, not a cherry-picked demo reel, which tells you how far generalist manipulation actually is.
Multi-robot collaboration is the new trick
Different robot types can now coordinate on a workflow neither finishes alone, with ER 2 orchestrating the hand-offs. It is the clearest signal yet that running a fleet of heterogeneous robots is becoming a software-orchestration problem, not just a hardware one.
Who's building on it
Launch partners span Apptronik's Apollo 2, Boston Dynamics, Franka Duo, and Agile Robots platforms. On-Device 2 claims it can adapt to a new robot body in hours from a small number of demonstrations, which is the piece that makes the models portable across hardware.
Elsewhere: GCC won't take your AI-written patches
The GCC steering committee will now decline any legally significant contribution containing LLM-generated code, effective immediately; test cases are the lone exception, and the policy is set to be revisited in early 2027. If you contribute to GCC, your Copilot- or Claude-assisted diffs won't be accepted, so keep AI-assisted work to trivial, clearly-marked changes.