Google launches Gemini Robotics ER 2 for real-time robot orchestration

Google DeepMind has launched Gemini Robotics ER 2, an embodied-reasoning model designed to serve as a high-level control layer for robots. Developers can use it to interpret the physical world, communicate with people, plan multi-step tasks, and delegate movement to lower-level vision-language-action models.

From observation to action

A key upgrade over Gemini Robotics ER 1.6 is continuous video understanding. Rather than treating each action as an isolated request, ER 2 can monitor a robot’s progress, recognize when something has gone wrong, adjust its plan, and decide when to continue. That makes the model better suited to environments where timing and changing conditions matter.

The system can also call external tools, including Google Search and developer-defined functions. Through the Gemini Live API’s bidirectional streaming, it is designed to coordinate actions with less delay between reasoning and execution. Google demonstrated the approach with Boston Dynamics’ Spot, using natural-language instructions and Spot APIs for navigation and manipulation to fetch an item.

Built for physical AI workflows

ER 2 supports multimodal video, audio, and text inputs, while developers expose robot-control interfaces—such as navigation APIs or VLA models—as callable tools. Google says the model outperforms ER 1.6 in tool orchestration across simulated and real robot-control settings, as well as human teleoperation.

The release also adds multi-robot collaboration, allowing multiple machines to share spaces and divide complex workflows. That could give AI builders a higher-level coordination layer instead of requiring every behavior to be hard-coded into an individual robot.

Gemini Robotics ER 2 is available through the Gemini API and Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform. Google has also published starter examples on GitHub.

Source: Google DeepMind Blog