Gemini Robotics
GoogleGemini Robotics is Google DeepMind's family of closed robot foundation models: Gemini Robotics 2, a vision-language-action model that turns vision and language into motor control; Gemini Robotics ER 2, an embodied-reasoning model for spatial planning offered through the Gemini API; and On-Device 2, a lightweight version that runs locally on the robot.
Openness
1 high confidence- weights
- closed(hosted API and trusted-tester access only)
- data
- closed
- code
- closed
- license
- proprietary(Gemini API terms)
Nothing downloadable: the action model reaches a trusted-tester program and the reasoning model the Gemini API, both under Google's service terms.
- https://ai.google.dev/gemini-api/docs/robotics-overview recorded 2026-09-27
Gemini API robotics overview: the ER model "is currently in preview" as gemini-robotics-er-2-preview; no weights, and requests need a restricted API key.
- https://deepmind.google/models/gemini-robotics/ recorded 2026-09-27
Product page: Gemini Robotics 2, ER 2 and On-Device 2 described; access through a waitlist, "100+ trusted testers"; no download, weights, training data or code link.
Adoption
3 low confidenceGoogle names more than a hundred trusted testers building on the action model, among them Agile Robots, Apptronik and Boston Dynamics, and the embodied-reasoning model is open to any Gemini API developer. No usage count is published.
- https://ai.google.dev/gemini-api/docs/robotics-overview recorded 2026-09-27
API docs: gemini-robotics-er-2-preview and a streaming preview available through the Gemini API.
- https://deepmind.google/models/gemini-robotics/ recorded 2026-09-27
Product page: "We're working with 100+ trusted testers — from enterprise automation leaders to cutting-edge robotics startups"; partners Agile Robots, Apptronik, Boston Dynamics.
Capability
5 low confidenceGemini Robotics covers more than any open policy here: one model family drives bimanual arms and full humanoid bodies, handles fine dexterous work, and pairs the action model with a separate reasoning model that plans and calls tools. The claims are the vendor's, with no public benchmark to check them against.
- https://ai.google.dev/gemini-api/docs/models/gemini-robotics-er-2-preview recorded 2026-09-27
ER 2 model card: text, image, video and audio input, function calling, code execution and thinking supported; latest update July 2026.
- https://deepmind.google/models/gemini-robotics/ recorded 2026-09-27
Product page: "Can be adapted to any bi-arm robot in just a few hours, scaling its intelligence from arms to complex humanoid bodies"; "screwing in a light bulb and tying knots"; "Controls entire humanoid bodies from feet to fingertips"; ER 2 "capable of reasoning within physical spaces to make detailed plans".
Verified 2026-09-27