Dexterity is still the weak point
The most honest part of the announcement is its own benchmark chart. Running an Apollo humanoid with five-fingered SharpaWave hands, Gemini Robotics 2 unscrews a lightbulb 92% of the time but screws one back in only 36% of the time. Tying a trash bag lands at 44%, a ziplock task at 40%, and a dustpan task at 32%. The spread between removing a bulb and reinserting it shows how much of fine manipulation remains unsolved once a task requires precise, reversible force rather than a grab-and-pull. DeepMind also concedes movement speed still needs work.
Two-finger grippers do better on their own benchmarks. On a Franka Duo, the model reports 74.2% on general pick and place, 78.9% on diverse tool kitting and 89.6% on precise insertion. The gap between gripper and multi-finger numbers is the practical story: the harder the hand, the further the reliability falls from anything deployable.
Transfer across robot bodies
Robotics has never scaled cleanly because robots differ in sensors, hands, joints and control stacks, so a policy learned on one body rarely survives on another. DeepMind's claim for the on-device model is that it adapts to a new bi-arm embodiment in a few hours, typically with fewer than 200 examples. If that holds outside the lab, it attacks one of the field's real operational load. Local inference matters here too: a robot moving near people cannot wait on a cloud round trip for every safe stop or navigation decision.
The ER model's other advance is duration. DeepMind says it can run task sequences lasting several minutes across hundreds of decisions, tracking when tasks begin and end, catching key events and self-correcting. Long-horizon reliability, not a single clean demo, is what separates a lab result from a working system.
Who can actually check this
Access is narrow, and that shapes how much the claims can be trusted. Gemini Robotics ER 2 is available in Google AI Studio and in private preview on the Gemini Enterprise Agent Platform. The vision-language-action model and the on-device model go only to early-access partners. Outsiders can experiment with the planning layer, but the models that translate perception directly into motor control are not open. Independent verification of the full-body, dexterity and transfer claims will be limited until that changes.
On safety, DeepMind introduces ASIMOV-Agentic, a benchmark for whether an embodied agent will refuse unsafe tool calls and detect human proximity, and says ER 2 scores better on safety-constraint following. These are the company's own numbers on the company's own benchmark. Because robotics failures can cause physical harm, that layer needs external scrutiny more than any other.
The signal worth watching is whether early-access partners reproduce these results on their own hardware and workflows, with failure rates and completion data on environments the demos never showed. If DeepMind keeps the direct motor-control models restricted, Gemini Robotics 2 will shape the market through controlled deployments rather than an open developer ecosystem.