Google DeepMind Releases Gemini Robotics-ER 2, a Model That Teaches Robots to Work Together

Google DeepMind announced the Gemini Robotics-ER 2 model on July 30, 2026, designed to serve as a "high-level brain" for robots. The novelty is that different machines, from the four-legged Spot to the humanoid Apollo 2, can now share tasks and inform each other of progress instead of operating in complete isolation.

What's new in ER 2

Gemini Robotics-ER 2 differs from its predecessors in that it plans its next moves while the robot is already moving, rather than during pauses to "think." The model analyzes camera footage in real time, monitors the progress of the task at hand, and corrects its own errors when something goes wrong.

Google DeepMind describes ER 2 as a qualitative leap in video understanding, task organization, and coordinating multiple robots at once. The model can also call external tools, such as Google Search or developer-defined functions, letting a robot pull in information beyond its own sensors.

How robots work together

In the demonstrations shown at launch, different types of robots divide tasks according to their own physical constraints. Wheeled robots handle work on flat indoor floors, while legged machines like Spot and Apollo 2 take over tasks on uneven terrain. A shared understanding of the situation across all the machines is meant to let them coordinate the sequence of actions without additional human oversight.

In one demonstration, Spot was connected to Boston Dynamics' own programming interface and carried out fetch-and-retrieve tasks, meant to show that the ER 2 model is suited for deployment on existing hardware, not just on Google's own prototypes.

Safety around people

Google highlights improved results in safety tests compared with ER 1.6, including adherence to safety instructions and response to human proximity. The model can independently stop a robot when a person comes near it, meant to reduce the risk of collisions when working in the same space as people.

The model's ability to read instrument readings has also been expanded, it now recognizes ten types of measuring devices, and its ability to answer questions about the robot's camera feed has been improved.

Significance for the industry

ER 2 is a planning-and-reasoning layer model, not one that controls a robot's body movement, that function belongs to a separate Gemini Robotics family. Google thus splits the tasks: one model handles what the robot should do and in what order, while another handles how it physically executes the movement.

For companies deploying fleets of robots in warehouses, factories, or logistics centers, this means the ability to combine hardware from different manufacturers under a single decision-making system, instead of programming separate coordination logic for each type of machine individually. Public access through the Gemini API also lowers the barrier to entry for smaller integrators who previously lacked the resources to build their own planning systems.

The model is reaching developers at a moment when competition in the market for large-language-model-based robotics is accelerating, with companies like Boston Dynamics and Apptronik increasingly integrating their hardware with external reasoning layers instead of building their own AI systems from scratch.