Thought Leadership
Google DeepMind has introduced Gemini Robotics 2, an AI model designed to control a wide range of robot types, including humanoid robots, with full body coordination.
Gemini Robotics 2 consists of multiple interconnected models working together. A vision language model interprets images and video, communicates with humans, and plans how a robot should execute complex, multi-step tasks. Two so called vision language action models then translate these plans into precise movements, controlling both the robot’s entire body and individual grippers or hands.
Unlike earlier DeepMind robotics models, which primarily focused on controlling the upper body for tabletop tasks, Gemini Robotics 2 can now operate the complete robotic system. This includes movements such as walking, crouching, stretching, and balancing in confined environments designed for humans.
In publicly released video demonstrations, Google DeepMind showcased several applications, including Apptronik’s humanoid robot Apollo 2, which used grippers from Sharpa to organize shelves. Other robots demonstrated tasks such as closing cable ties and replacing a light bulb. According to Google DeepMind, the models were trained using a combination of human remote operation, video examples, and simulations. However, the system does not yet have a general ability to perform any new task without task specific training.
New Safety Benchmark for Collaborative AI Systems
Google says it is pursuing a multi layer safety strategy, with dedicated protection mechanisms integrated at every model level. As part of this approach, the company introduced ASIMOV Agentic, a new benchmark designed to evaluate the safety of multiple interacting AI systems used for robotic control.
The benchmark examines whether a higher level reasoning model rejects unsafe commands before they are passed to the action control model. It also evaluates whether the system can identify when a task is impossible to complete or requires human intervention.
According to Google DeepMind, the perception focused model Gemini Robotics ER 2 is among the company’s safest robotics models so far when it comes to maintaining safe distances from nearby humans.
For the development of Gemini Robotics 2, Google DeepMind is collaborating with robot manufacturers including Apptronik, Boston Dynamics, and Agile Robots, as well as more than 100 selected testing partners.
(ll)
