Gemini Robotics ER 2 lands in the API: one brain for many robots
A vision-language model bringing spatial reasoning, video understanding and multi-robot orchestration to one Gemini endpoint — old ER 1.6 preview shuts Aug 31.
Copy markdown
A high-level brain you reach over HTTP
Gemini Robotics ER 2 is a vision-language model for embodied reasoning, now callable as gemini-robotics-er-2-preview (plus a streaming variant) in the Gemini API and AI Studio. It takes text, image, video and audio, does spatial reasoning and video understanding, and can orchestrate multi-step tasks across more than one robot.
Function-call your hardware
You declare low-level controllers — including vision-language-action models — as tools and stream multimodal data straight in, so the model plans while your actuators execute. It ships with a 131K-token input / 65K-token output window and supports code execution, function calling, structured outputs and search grounding.
The benchmark numbers
DeepMind reports 91.3% accuracy finding critical moments in a feed (0.96s mean error) and 57.4% classifying task progress across five levels, while running about 4x faster than larger tiers at sub-second latency. Demos drove Boston Dynamics' Spot and paired Apptronik's Apollo with a Franka arm on shared tasks.
Migrate off ER 1.6 before Aug 31
Heads-up if you are already on the preview line: gemini-robotics-er-1.6-preview is scheduled to shut down on August 31, so move any embodied-reasoning workloads over to ER 2 before then.
Elsewhere: Gemini's Flash tier drops temperature and top_p
In the same changelog, Google deprecated the temperature, top_p and top_k sampling parameters on gemini-3.6-flash and gemini-3.5-flash-lite. Calls that set them will quietly stop taking effect, so audit any Flash paths that lean on custom sampling.