Google DeepMind has unveiled Gemini Robotics ER 2, a significant upgrade to its embodied reasoning model for robotics. Designed as a high-level brain for robots, the model enhances video understanding, task orchestration, and multi-robot collaboration, enabling robots to operate more effectively in the physical world.
Gemini Robotics ER 2 allows robots to process continuous video feeds to track their own progress, adapt to errors in real time, and know precisely when to move to the next step. It can hand off motor execution to lower-level vision-language-action (VLA) models and natively call tools like Google Search or user-defined functions. The model is designed to think about next steps while simultaneously performing actions, reducing the jarring 'stop-and-think' pauses common in earlier systems.
A key advancement is multi-robot collaboration, where different robots—such as a wheeled rover and a humanoid—can communicate via shared semantic understanding to complete complex tasks that a single robot cannot handle alone. Google demonstrated this with Apptronik's Apollo 2 and Franka F3 Duo working together.
The model also brings improvements in temporal intelligence, including progress classification and moment finding. On progress classification, Gemini Robotics ER 2 achieves 57.4% accuracy, outperforming previous models. For moment finding, it achieves 91.3% accuracy with a mean absolute distance of 0.96 seconds, delivering precision at a fraction of the compute cost and four times the execution speed of larger models.
In spatial reasoning, the model excels at success/failure detection from raw video, general instrument reading across 10 instrument types, and enhanced visual question answering. It also sets new safety benchmarks, outperforming its predecessor on Safety Instruction Following and Human Proximity metrics, and can autonomously halt a humanoid robot when a person is nearby.
Gemini Robotics ER 2 is now publicly available to developers via the Gemini API, Google AI Studio, and in private preview on the Gemini Enterprise Agent Platform. Google has released example code on GitHub to help developers get started.