- Published on
Gemini Robotics ER 2 Debuts, Supporting Continuous Video Understanding and Multi-Robot Collaboration
- Authors
- Name
- DeepHorus
On July 30, Google's DeepMind released the next-generation robotic embodied reasoning model, Gemini Robotics ER 2, positioned as the 'advanced brain' for robots. The model can coordinate human-robot interaction, understand the physical world, and plan multi-step tasks, while assigning specific tasks to lower-level vision-language-action models. Compared to the previous version 1.6, the biggest breakthrough of ER 2 lies in its ability to analyze continuous video streams. Robots can use continuous video input to track their own progress in real time, automatically adjust or retry when actions go wrong, and accurately determine when to move to the next step, thus changing the traditional 'stop-think-act' loop. In tests, ER 2 achieved 57.4% accuracy in task progress classification, and could correct failed steps without restarting the entire workflow. In critical moment localization tests, the model achieved 91.3% accuracy with an average absolute time error of only 0.96 seconds, for example, accurately determining when to stop pouring coffee into a cup. Google stated that ER 2 achieves accuracy close to that of larger models at lower computational cost, with a 4x improvement in execution speed, and meets the sub-second latency requirements of real-world robot operations. In addition, the model natively supports calling Google Search or developer-defined functions, further expanding the robot's capabilities in information retrieval and tool use. ER 2 also supports multi-robot collaboration: different types of robots can communicate and hand off tasks through shared semantic understanding, completing complex workflows that a single device would struggle to handle independently. Google demonstrated a collaboration case between Apptronik's Apollo 2 and Franka's FR3 Duo to illustrate the practical application of this capability. For physical agents, ER 2 can synchronously reason about subsequent steps while the robot executes tasks, reducing latency. The model has been integrated into the bidirectional streaming endpoint of the Gemini Live API, making it particularly suitable for latency-sensitive robot tasks and helping improve the response speed and flexibility of robots in dynamic environments. Currently, Gemini Robotics ER 2 is available to developers through the Gemini API and Google AI Studio, and has also joined the private preview of the Gemini Enterprise Agent Platform. Google also released model configuration and prompt example code to help developers get started quickly. This marks an important step toward more efficient and collaborative robotic embodied reasoning technology.
