IllustrationNo official image on file. Showing a schematic category illustration.
Embodied-reasoning model that plans multi-minute tasks, talks with people and coordinates multiple robots, handing off execution to a VLA.
Key specs
| Product type | Embodied-reasoning VLM |
|---|---|
| Availability | Google AI Studio; private preview on Gemini Enterprise Agent Platform |
Technical details
Gemini Robotics ER 2 is DeepMind's embodied-reasoning VLM that acts as a robot's high-level agent: it perceives the scene, plans multi-step tasks lasting several minutes, calls a VLA or custom robot functions, and tracks progress and success. New in ER 2 are video moment finding and progress classification, multi-robot orchestration and multi-step tool use, plus a low-latency streaming endpoint for continuous audio/video. It accepts text, image, video and audio input with a 131,072-token input limit and returns text or structured spatial output. DeepMind also introduced the ASIMOV-Agentic safety benchmark alongside it.
| Architecture | Embodied-reasoning vision-language model that builds on Gemini 3.5 Flash; offered as gemini-robotics-er-2-preview and a Live-API streaming variant gemini-robotics-er-2-streaming-preview. |
|---|---|
| Inputs outputs | Inputs: text, images, video, audio (131,072-token input limit); output: text, incl. structured points/boxes/trajectories and function calls (65,536-token output limit). |
| Embodiments supported | Robot-agnostic orchestrator: calls VLAs (e.g. Gemini Robotics 2) or user-defined robot functions; supports multi-robot coordination. |
| License weights | Proprietary; API access only. |
| Hardware requirements | Cloud API (Gemini Developer API / Google AI Studio); private preview on Gemini Enterprise Agent Platform. |
| How to try | ai.google.dev/…/robotics-overview |
| Sources |
