Google DeepMind debuts Gemini Robotics 2 model series for humanoid robots

Alphabet Inc.'s artificial intelligence research lab, Google DeepMind, unveiled a new family of models called Gemini Robotics 2, designed specifically for humanoid robots. The innovative series of models enables multiple autonomous robots to collaborate on a single task, automating chores that typically require hundreds of steps.

A key distinction of the Gemini Robotics 2 series lies in its "embodied reasoning algorithm" known as Gemini Robotics ER 2. Unlike previous models, ER 2 can comprehend tasks described in natural language, making it more accessible for users. This algorithm can manage lengthy tasks by dividing them among several robots, enhancing efficiency.

To address the need for robots to repeat tasks they fail initially, DeepMind incorporated two features into ER 2. Firstly, it can track a task's progress using footage from the robot's cameras. If a mistake is made, ER 2 can pinpoint the last successful step and continue from there, saving time. Secondly, ER 2 can better determine when a task is complete, helping the robot move on to the next task promptly and facilitate problem-solving.

After ER 2 generates a plan for a task, it is sent to one of two other models in the Gemini Robotics 2 series - Gemini Robotics 2 and Gemini Robotics On-Device 2. Gemini Robotics 2 can control all components of a humanoid robot, optimizing its center of gravity to minimize the risk of falls and supporting a wider range of robotic hands. Gemini Robotics On-Device 2, on the other hand, runs directly on the robot's onboard computer and can be adapted to new robots with minimal training.

Google Cloud, the Gemini API, and Google AI Studio provide developers access to ER 2. The company also introduced the ASIMOV-Agentic Benchmark, a new safety benchmark designed to evaluate robots' ability to avoid collisions and other risks. The rollout of the Gemini Robotics 2 series comes alongside this new safety benchmark.