Google DeepMind releases Gemini Robotics 2: General-purpose AI "brain" adapts to new body within hours
Google DeepMind has released the Gemini Robotics 2 series of general-purpose AI "brain" smart kits. The three models are adapted to different hardware bodies across models. They can be adapted to a new body after a few hours of data training, and the robot's "smart layer" is positioned for development by hardware manufacturers.
Google DeepMind released the Gemini Robotics 2 series, positioning it as a general-purpose AI "brain" intelligence suite, providing a unified intelligence layer for embodied intelligent robots. The three models can be adapted to different hardware bodies across models. The official adaptation speed is: it only takes a few hours of data training to connect the model to a brand new body.
Position the "intelligent layer" and decouple it from the hardware
DeepMind clearly does not build the entire machine. Instead, it uses Gemini Robotics 2 as the "intelligent layer" output of the robot, which is responsible for core capabilities such as perception, understanding, planning, and execution, allowing hardware manufacturers to focus on the design of the entire machine and the implementation of scenarios on this basis. This approach is exactly the same as the "Android system in the robotics world" - whoever masters the general intelligence layer controls the entrance to the embodied intelligence ecosystem.
Hours of adaptation: software redefines hardware rhythm
Traditional robot skill development takes months, but Gemini Robotics 2 compresses the adaptation of new models to hours. This change will completely rewrite the rhythmic relationship between hardware iteration and software intelligence: hardware manufacturers can significantly reduce investment in AI R&D and concentrate resources on structures and scenarios; however, the generalization performance of truly complex industrial scenarios still needs to wait for actual implementation data to be tested.
Layered combination of three models
The three models in the series cover different computing power and form requirements. Through the layering of large cloud models + small end-side models, gradient divisions are made between cost, latency and capabilities to provide corresponding solutions for robots at different price points. This is also the only path for embodied intelligence to move from "single point demonstration" to "volume production".
For domestic robot manufacturers, the adoption of Gemini intelligent layer means supply chain and data security considerations; the speed of follow-up of domestic intelligent layer solutions will determine the direction of this "intelligent layer standard" dispute, and will also directly affect the bargaining space of Chinese manufacturers in the embodied intelligence ecosystem.
Several directions worth tracking in the future:
- Model details: Specific parameters, form adaptation range and authorization/charging modes of the three models.
- Domestic Access: Whether to connect to domestic mainstream robot manufacturers, and how to follow up on domestic intelligent layer solutions.
- Industrial Generalization: Real generalization performance of "several hours of adaptation" in complex industrial scenarios.
- Standards battle: The evolution of the standard battle for embodied intelligence "smart layer" (Google vs. domestic solutions).
Reviews