Gemini Robotics 2 controls a full humanoid, feet to fingertips

Google DeepMind released three robotics models on July 30. The headline model, Gemini Robotics 2, is a vision-language-action model that drives a full humanoid from feet to fingertips, and it also runs bi-arm robots with five-fingered hands that have 22 degrees of freedom. Alongside it, Gemini Robotics ER 2 acts as the high-level planner, holding a task together across several minutes and hundreds of decisions, and it can coordinate more than one robot at once. A third model, Gemini Robotics On-Device 2, runs locally with no network and adapts to a new bi-arm robot from a few hours of data, usually fewer than 200 examples.

The success rates are worth reading closely, because they show what is solved and what is not. On an Apollo humanoid, picking from a shelf worked 76.3% of the time and from a table 68.4%, but picking off the floor dropped to 45.7%. Fine dexterity is more uneven: unscrewing a bulb hit 92%, yet screwing one back in reached only 36%, and sweeping with a dustpan sat at 32%. On a Franka gripper setup, precise insertion was strong at 89.6%. DeepMind also introduced a safety benchmark, ASIMOV-Agentic, that tests whether the planner refuses unsafe tool calls, predicts when a task is infeasible, and asks a human for help. You can read DeepMind's post.

Why it matters

If you build on robot hardware, an on-device model that adapts from under 200 examples changes how much data you need to try a new embodiment, and the honest per-task numbers tell you which jobs are ready to pilot and which still need a person standing by.