Google DeepMind has officially announced the release of a new generation of artificial intelligence for robotics — the Gemini Robotics 2 family of models. This update marks a transition from simple automated movements to complex cognitive processes in the physical world. Developers refer to this technology as 'Physical AGI' (Artificial General Intelligence), the main goal of which is to create a universal robot capable of performing a wide range of household and industrial tasks.
'Embodied Reasoning' and Real-Time Operation
The key innovation is the Gemini Robotics ER 2 model, which handles the 'embodied reasoning' function. The algorithm is capable of analyzing video streams from the robot's cameras in real-time, integrating with the Gemini Live API. This allows machines to track the progress of tasks step-by-step and independently correct errors without the need to restart the entire process.
According to developers, if a robot accidentally drops an object, the system will instantly adjust the movement rather than starting the action over. The accuracy of determining key moments of action has reached nearly 90 percent. This allows the robot to understand nuances, for example, precisely determining when to stop pouring coffee into a cup to avoid spilling the liquid.
Enhanced Coordination and Teamwork
The Gemini Robotics 2 model transforms text and visual commands into direct movements of mechanical limbs. Significant progress has been made in controlling complex humanoid hands with multiple fingers: the new version of the system ensures unprecedented accuracy and coordination.
Special attention has been paid to the robots' ability to interact. During testing, the Apollo 2 and Franka F3 Duo models demonstrated the ability to perform tasks jointly, without interfering with each other and effectively distributing roles within the team.
Autonomy and Rapid Adaptation
To ensure operation in conditions without internet access, a lightweight version was introduced — Gemini Robotics On-Device 2. It is capable of functioning locally, directly on the device, without connecting to cloud servers.
The system features high learning speed. The algorithm adapts to a new robot design in just a few hours of training or based on the analysis of only 200 examples. Currently, the technologies are being tested on equipment from Boston Dynamics.
Safety and Protection Against Hallucinations
Developers have addressed risks associated with AI errors and implemented special levels of protection against 'hallucinations' and false actions. Along with the release, a new safety benchmark named ASIMOV-Agentic was presented. The ER 2 model demonstrates a high ability to recognize potential danger and stop movement if a person approaches the working area too closely.
Along with the technical specifications, Google has published the pricing policy for the new development, paving the way for the commercial implementation of these technologies.
