Google DeepMind has unveiled Gemini Robotics 2, an artificial intelligence system designed to control a variety of robotic designs, from industrial arms to humanoids. The solution combines a task-planning model with models that translate images, language and commands into movement. The robot can walk, crouch, maintain its balance, manipulate objects and collaborate with other machines.
In demonstrations, Apptronik’s humanoid robot, Apollo 2, tidied shelves, moved objects and performed tasks requiring the use of its five-fingered hands. However, the results show that the technology is still at the development stage. The success rate for picking up objects from the floor was 45.7 per cent, for screwing in a light bulb 36 per cent, and for unscrewing it 92 per cent. The model is currently being made available to selected partners on a trial basis.
The launch is part of the rapid development of automation. According to the International Federation of Robotics, 542,000 industrial robots were installed worldwide in 2024, whilst sales of professional service robots rose by 9 per cent to nearly 200,000 units. Demand is being driven, amongst other things, by labour shortages and the need to boost productivity.
If similar models can be transferred relatively easily between robots from different manufacturers, companies could reduce the cost of programming each machine individually. Google, meanwhile, would have the opportunity to create a common software platform for robotics, similar to the role Android plays in the mobile device market.
However, this does not mean that humanoids will soon be entering offices and homes. The most significant barriers remain reliability, the cost of the hardware, liability for errors, and safety when working alongside humans. DeepMind addresses some of these issues with the ASIMOV-Agentic benchmark, which tests, amongst other things, the rejection of dangerous commands and the recognition of situations requiring human intervention.
