Google Unveils Gemini Robotics-ER 2 to Advance AI-Powered Robot Reasoning

Google DeepMind has introduced Gemini Robotics-ER 2, the latest version of its embodied reasoning model for robotics, expanding the company’s effort to bring Gemini-powered intelligence into physical machines. Announced on July 30, 2026, the new model focuses on improving how robots perceive their surroundings, reason through long-running tasks, coordinate with other robots, and safely interact with people.

The announcement marks Google’s latest step toward general-purpose robotics by separating high-level reasoning from low-level robot control, enabling robots to tackle more sophisticated physical tasks across different hardware platforms.

Gemini Robotics-ER 2 is an upgraded reasoning model that helps robots understand the physical world before deciding what actions should be taken. Unlike motion-control models that directly generate motor commands, ER 2 acts as a planning and decision-making layer capable of interpreting visual information, natural language instructions, and environmental context.

Google says the model can:

  • Understand complex physical environments

  • Plan multi-step tasks lasting several minutes

  • Coordinate multiple robots working together

  • Generate structured plans that action models or robotics APIs can execute

  • Adapt to different robot embodiments

The company introduced ER 2 alongside the broader Gemini Robotics 2 family, which also includes a vision-language-action model for full-body robot control and an on-device model designed to operate without cloud connectivity.

Moving Beyond Simple Robot Commands

Traditional industrial robots generally execute predefined sequences of instructions. Gemini Robotics-ER 2 aims to address a different challenge: enabling robots to understand goals instead of simply following scripts.

According to Google DeepMind, the model continuously processes information from cameras, language inputs, video streams, and other sensors before generating plans for completing a task. Rather than controlling individual motors directly, it determines what should happen next and delegates execution to robot-specific control systems.

This architecture allows the same reasoning engine to work across multiple robot designs without requiring identical hardware.

Longer Tasks and Multi-Robot Collaboration

One of the most notable additions is support for long-duration planning.

Instead of responding to a single instruction at a time, Gemini Robotics-ER 2 can manage workflows that involve numerous intermediate decisions. The company demonstrated scenarios where robots complete multi-step activities while adapting to changing conditions.

Google also introduced multi-robot coordination capabilities. In these scenarios, several robots can divide responsibilities, share progress, and complete tasks collaboratively rather than operating independently.

Such capabilities could prove valuable in manufacturing, warehouses, logistics facilities, and large industrial environments where multiple autonomous systems already operate simultaneously.

Improved Embodied Reasoning

Embodied reasoning refers to an AI system’s ability to connect digital intelligence with physical interaction.

Gemini Robotics-ER 2 builds on previous research by improving spatial understanding, object recognition, environmental awareness, and task planning. The model is designed to interpret scenes more accurately before generating plans for robotic execution.

Google says the reasoning model can:

Better understand physical environments

The model analyzes objects, locations, and spatial relationships before planning movements.

Generate complex task plans

Rather than issuing isolated commands, ER 2 creates structured workflows that guide robots through complete jobs.

Support multiple robot types

Because planning is separated from movement control, developers can integrate the reasoning system with different robot platforms.

Safety Remains a Core Focus

As robotics systems become increasingly autonomous, safety becomes equally important.

Google says Gemini Robotics-ER 2 includes several mechanisms intended to reduce operational risk, including:

  • Detecting nearby people

  • Triggering safety stops when necessary

  • Refusing unsafe requests

  • Asking for human assistance when uncertainty exceeds acceptable limits

The company has also expanded its robotics safety evaluation efforts to assess how AI-powered robots behave in real-world situations before broader deployment.

Availability

Google is taking a staged rollout approach.

Gemini Robotics-ER 2 is available through Google AI Studio for developers experimenting with embodied reasoning. Meanwhile, Gemini Robotics 2 and the on-device robotics model are initially being released through an early-access program for selected partners.

Google has not announced a general public release timeline for the complete robotics platform.

No pricing information was disclosed.

Conclusion

With Gemini Robotics-ER 2, Google DeepMind continues expanding Gemini beyond digital assistants and software into physical AI systems capable of understanding and interacting with the real world. Rather than concentrating solely on robot movement, the company is investing in the reasoning layer that allows machines to interpret complex environments, coordinate actions, and complete longer tasks more autonomously.

The next phase to watch will be broader developer adoption, additional hardware partnerships, and evidence that these foundation models can perform reliably outside controlled demonstrations.