Google DeepMind introduced the Gemini Robotics 2 model family on July 30, 2026, three models that the lab says give robots “intelligent whole-body control, fine dexterity, and teamwork.” The release covers Gemini Robotics 2, a vision-language-action (VLA) model; Gemini Robotics ER 2, an embodied-reasoning model; and Gemini Robotics On-Device 2, an efficient VLA that runs on the robot itself.
On this page
What the three models do
| Model | Role, per Google DeepMind | Availability |
|---|---|---|
| Gemini Robotics 2 | VLA model that turns instructions into robot actions | Early-access partners |
| Gemini Robotics ER 2 | “High-level brain”: planning, spatial reasoning, progress tracking | Gemini API, Google AI Studio; private preview on Gemini Enterprise Agent Platform |
| Gemini Robotics On-Device 2 | Efficient VLA for on-device use | Early-access partners |
According to Google DeepMind, Gemini Robotics 2 can control “full humanoids, from feet to fingertips, and other bi-arm robots.” In its examples, a humanoid has to “walk, crouch, stretch, and manipulate objects to clean up a cluttered room,” and follows instructions such as “put the watering can into the green bin in the bottom shelf.”
The same model works across different hands and grippers. Google DeepMind names a “five-fingered, 22 degree-of-freedom SharpaWave hand,” “standard two-fingered parallel grippers,” and a two-fingered gripper on a Franka Duo platform. Robots shown in the announcement include Apptronik’s Apollo 2 humanoid.
ER 2 as the planner
In the Gemini Robotics 2 stack, ER 2 “serves as the robot’s high-level brain,” Google DeepMind writes: it processes user instructions, observes the room, reasons about the steps, coordinates with the VLA model and tracks progress until the task is done. Google says ER 2 can watch continuous video feeds so robots can “track their own progress, adapt if something goes wrong.”
The release also adds multi-robot collaboration. Google DeepMind says it lets “different types of robots to communicate and work together to solve complex workflows a single robot could not do alone.” The ER 2 post shows Boston Dynamics’ Spot, Apptronik’s Apollo 2 and Franka’s F3 Duo among the robots used.
Results reported by Google
Google reports two video-understanding results for ER 2 from its own evaluations:
| Task (Google’s evaluation) | ER 2 result |
|---|---|
| Progress classification | 57.4% accuracy |
| Moment-finding | 91.3% accuracy, 0.96 s mean absolute distance |
Partners and access
Google DeepMind thanks the Apptronik, Boston Dynamics and Agile Robots teams for their support. Developers can use ER 2 now through the Gemini API and Google AI Studio, while the Gemini Robotics 2 VLA and On-Device 2 models are limited to early-access partners. For other Google model releases, see the AI model release timeline.
