On 30 July 2026, Google DeepMind released Gemini Robotics ER 2 — an embodied reasoning model that acts as a “high-level brain" for robots. It plans multi-step tasks, analyzes a video stream in real time and coordinates multiple robots, while delegating the actual motion to separate action models. It is the thinking layer that pairs with Gemini Robotics 2, the body-control model unveiled a day later.
Key takeaways
- An embodied reasoning model serving as a “high-level brain" for robots
- Real-time video analysis: 57.4% progress classification, 91.3% moment-finding (0.96s mean error)
- Claimed 4x execution speed versus larger models
- Tool orchestration: calling VLA models, APIs and Google Search
- Available via Gemini API, Google AI Studio and a private preview on Gemini Enterprise Agent Platform
Division of labor: thinking apart, motion apart
Gemini Robotics ER 2 does not drive motors directly. The model handles spatial reasoning and planning, and delegates motion to VLA (vision-language-action) models. Separating the “brain" layer from the “muscle" layer is the key move — reasoning can run slower and deeper, while the fast control loop stays with a specialized action model.
The model treats video as a continuous stream, not a series of separate frames. That lets it track task progress, detect success or failure, and self-correct when something goes wrong. It can also read instrument displays and answer questions about the space around the robot.
Results: speed over size
The most telling number is not a single benchmark but speed. Google cites 4x faster execution than larger models, while keeping reasoning quality. On video, the model reaches 91.3% accuracy in finding the right moment (with a 0.96-second mean error) and 57.4% in classifying progress across five levels. These results matter for a robot that must know which stage of a task it is in before making its next move.
| Video task | Result |
|---|---|
| Progress classification (5 levels) | 57.4% |
| Moment-finding | 91.3% (0.96s mean error) |
ER 2 can also orchestrate tools — call a VLA model for motion, reach for an API or Google Search for information. That sets it apart from its predecessor, Gemini Robotics-ER 1.6, and shifts the focus from a single task to coordinating many steps and many robots at once.
Where and how to use it
The model is publicly available through the Gemini API and Google AI Studio, and in a private preview via the Gemini Enterprise Agent Platform. DeepMind also published code examples on GitHub. The authors are engineers Steven Hansen and Peng Xu.
Why it matters
Separating reasoning from control is an architectural bet on how general-purpose robots will be built. Instead of one giant model doing everything, there is a “brain" layer that plans and corrects, plus interchangeable action models that execute motion. That split makes scaling easier, since each layer can evolve independently and a single reasoning model can drive different robots. The emphasis on speed over sheer size suggests the bottleneck in robotics is no longer the model's knowledge, but how fast it can react to a changing world.
What's next
- The Gemini Robotics 2 action model (whole-body control), announced on 31 July, adds the body-control layer to ER 2 — together they form a full stack
- ER 2 is already available in the Gemini API and Google AI Studio, so developers can test multi-robot orchestration right away
- The private preview on the Gemini Enterprise Agent Platform points to the next step: commercial deployments in industrial settings
Sources
- Google — Introducing Gemini Robotics ER 2
- IEEE Spectrum — Video Friday: Meet Google DeepMind's Gemini Robotics 2
