At a time when humanoid robots are accelerating their deployment, the iteration speed of AI "brains" has once again refreshed industry perceptions. A relevant team has recently officially launched a new embodied reasoning model - Gemini Robotics ER2. As a core driving system specifically designed for hardware devices, this model not only advances in physical environment understanding and multi-step task planning, but also successfully solves the challenge of multi-robot collaboration at the technical level for the first time.

Since the introduction of the related embodied intelligence system, the development team has continuously accelerated product iterations. After the evolution of previous versions, the new ER2 model has made breakthroughs in real-time decision-making and continuous task monitoring. Unlike traditional reasoning models that require "pausing to think" before executing actions, this model allows robots to plan their next steps while maintaining movement.

image.png

Through deep integration with real-time voice interaction interfaces, hardware devices can process continuous tasks more smoothly, significantly reducing intermediate stutters and pauses. At the same time, the new model has the ability to continuously understand video content, not only evaluating task progress in real-time, capturing error-prone steps during execution, but also directly retrying problematic steps, eliminating the cumbersome process of resetting the entire workflow. Test data shows that the model demonstrates extremely high accuracy in task progress classification and key event timing identification.

Notably, the model has officially introduced a multi-robot collaboration mechanism. Relying on this technical foundation, different forms and functions of robots can share environmental data and efficiently achieve division of labor handoffs. In an official test demonstration, the relevant humanoid robots and robotic arm platforms successfully collaborated, vividly demonstrating the capabilities of cross-device collaborative operations.

Currently, the model is open to the industry through APIs and developer platforms, and is also undergoing private testing on enterprise-level Agent platforms, marking a solid step forward for embodied intelligence toward large-scale commercialization.