Google DeepMind’s Gemini Robotics 2 arrived at the end of July with videos of a humanoid tidying a room and language about robots “of any shape or size” and a milestone “on the path toward solving AGI in the physical world.” The same release publishes, chart by chart, how often the robots actually succeed, and says who can use each model. Read in that order, it is a laboratory advance with partial access.
Three models, three levels of access
The release is three models. Gemini Robotics 2 is a vision-language-action model that turns instructions into motor control. Gemini Robotics ER 2 is an embodied-reasoning model that plans and orchestrates, built on Gemini 3.5 Flash. On-Device 2 is the most efficient action model, designed to run locally on the robot.
Only ER 2 is open to developers, in public preview on Google AI Studio and the Gemini API as gemini-robotics-er-2-preview, with a streaming variant for the Live API. The action model is in private preview for early-access partners, and On-Device 2 is limited to trusted testers. Neither has a public endpoint or price, and no weights for any of the three were found on Hugging Face.
ER 2 is free on the API’s free tier, where content is used to improve Google’s products. On the paid tier it costs $1.00 per million input tokens and $5.00 per million output tokens through December 31, 2026, and output rises to $10.00 from January 1, 2027, according to the pricing page. ER 1.6, its predecessor, is being shut down.
What the success rates say
The action model’s performance charts show three levels:
- Whole-body tasks on an Apptronik Apollo 2 humanoid: picking up objects from a table 68.4% of the time, from the floor 45.7% and from a shelf 76.3%.
- Gripper tasks on a Franka dual-arm system: general pick and place 74.2%, tool kitting 78.9% and precise insertion 89.6%.
- Multi-finger tasks with a dexterous hand: unscrewing a bulb 92%, but screwing it in 36%, tying a trash bag 44%, using a dustpan 32% and closing a ziplock bag 40%.
Most bars average several tasks within a skill category; the multi-finger chart shows each task. Google says it plainly: while the model reaches “a medium to high success rate” on whole-body and gripper tasks, “the multi-finger dexterous manipulation remains challenging.”
On-Device 2 is described as adapting to new two-arm robots in a few hours with fewer than 200 examples, but no success rates for those platforms appear on the pages reviewed.
Safety as orchestration
Google’s answer to the risk of a robot executing an unsafe or impossible action is to use ER 2 as a safety layer: refusing unsafe tool calls, watching human proximity and asking for help when uncertain. On Google’s own benchmarks, ER 2 scores 97.9% on following safety instructions and 93.0% on detecting a person within one meter. Its safety report adds a caveat to those figures: models that classify a constraint correctly in text are less reliable when they have to turn it into coordinates or an executable action. Google also released ASIMOV-Agentic, a benchmark for this kind of safety orchestration, as an open dataset.
The company adds that it has not tested the models on every kind of robot and that running them safely is the user’s responsibility.
Related reading
- Gemini Robotics ER 2 in the Gemini API — Google AI for Developers
- Gemini Robotics 2 safety report — Google DeepMind
- NVIDIA Cosmos 3: The First Open Model That Understands the Physical World — IA al Día