Google DeepMind on Wednesday released Gemini Robotics 2, a three-model suite that gives humanoid robots coordinated whole-body motor control for the first time and lets a single model checkpoint run across different hardware bodies. The launch reframes what had been a hands-and-eyes problem as a systems problem, and it lands in a market where the pitch decks have started to precede the physics.
The suite splits into a vision-language-action model (Gemini Robotics 2), a higher-level embodied-reasoning model that can orchestrate multiple machines on a single task (Gemini Robotics ER 2), and a local variant, Gemini Robotics On-Device 2, that DeepMind says can adapt to entirely new robot embodiments with “a few hours” of data. In the launch demo, an Apptronik Apollo 2 executed a multi-step pick-and-place from a single natural-language prompt: “put the watering can into the green bin in the bottom shelf.” The same checkpoint, DeepMind says, drives Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and a Franka Duo with a Robotiq gripper.
The dexterity numbers are the tell. Bloomberg reports a 92% success rate on unscrewing a light bulb. The Next Web, citing figures from the Chosun Daily, puts the trash-bag tie task at 44% and the ziplock seal at 40%. Kanishka Rao, DeepMind’s robotics director, concedes that true dexterity is still distant and that robots learn far less efficiently than humans do. That’s the honest read: whole-body coordination is now legible to the model, but fine manipulation still isn’t.
DeepMind is packaging the release with safety scaffolding. ER 2 is billed as its safest reasoning model yet, better at noticing when a person is nearby and pausing, and the team is publishing a new benchmark, ASIMOV-Agentic, to score refusal of unsafe commands. Engineers Steven Hansen and Peng Xu say ER 2 is available through the Gemini API, Google AI Studio, and a private preview on the Gemini Enterprise Agent Platform. More than 100 trusted testers have access. Named hardware partners include Apptronik, Boston Dynamics, and Agile Robots.
Carolina Parada, DeepMind’s head of robotics, told Wired the release is “another milestone in our path towards really getting towards what we call like physical AGI.” That phrase is doing narrative-management work. OpenAI and Nvidia are building their own robot models, and Elon Musk continues to pitch Optimus in $10 trillion terms. Against that backdrop, DeepMind is claiming the substrate layer, the model that runs on everyone else’s metal, and quietly conceding the ziplock still wins four times out of ten.
The interesting scoreboard isn’t dexterity. It’s who owns the checkpoint.
Sources
- https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/
- https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/
- https://www.bloomberg.com/news/articles/2026-07-30/google-unveils-gemini-ai-for-robots-struggling-with-dexterity
- https://thenextweb.com/news/gemini-robotics-2-whole-body-humanoid-control
- https://www.engadget.com/2227268/google-gemini-robotics-2-platform-intelligent-whole-body-control/