Google DeepMind's Gemini Robotics 2 Gives Robots Whole-Body Control
Three new models handle motor control, planning, and on-device execution, and can adapt to a new robot body in a few hours instead of a full retrain.
DeepMind's new robot models learn a new robot body in hours, not a full retrain from scratch.
Google DeepMind released three robotics models on July 30, 2026: Gemini Robotics 2 for motor control, ER 2 for embodied reasoning and planning, and On-Device 2 for local execution without a network connection.
What the models actually do
DeepMind's benchmarks claim whole-body control, walking, crouching, dexterous manipulation with hands that have 22 degrees of freedom, multi-step task planning across minutes and hundreds of decisions, and multi-robot collaboration. Reported success rates run 45.7 to 76.3 percent on general whole-body manipulation tasks and 74.2 to 89.6 percent on gripper dexterity tasks.
The part that's actually new: fast adaptation
DeepMind says a robot running these models can adapt to a new hardware body, different arms, different grippers, in a few hours using motion-transfer techniques instead of a full retrain per robot. That's the practical claim worth watching, since it's what would let one model move across a fleet of different physical robots instead of needing a bespoke model per machine.
Who gets access first
This is early access, not general availability. ER 2 is live in Google AI Studio plus a private preview on the Gemini Enterprise Agent Platform. The motor-control and on-device models are restricted to early-access hardware partners, including Apptronik, Franka, Dexmate, and Trossen, by application. DeepMind's robotics lead, Carolina Parada, has been candid in interviews that deployment still carries real uncertainty, and that the safety question needs to be understood more deeply before robots operate with less supervision.
Why a build studio cares
Most of our work is web and software, not robotics, but the vision-language-action pattern here, one model turning messy real-world input directly into action, is the same architecture behind the browser and computer-use agents we do build with. Worth watching how DeepMind handles the safety guardrail layer, a new benchmark called ASIMOV-Agentic that checks whether an agent refuses an unsafe action and asks for help, since that same problem shows up in software agents with real tool access.
Next step: read DeepMind's announcement or Engadget's hands-on coverage. If agentic tooling is part of your roadmap, write to us at hello@gattyworks.com.