Google DeepMind has introduced Gemini Robotics 2, a new generation of artificial intelligence models built to control robots ranging from tabletop arms to full humanoids. The release moves the company’s robotics work beyond isolated picking and placing, adding whole-body movement, detailed physical reasoning and cooperation between multiple machines.
At the centre of the system is a vision-language-action model. It turns visual information and spoken or written instructions into physical actions, allowing a robot to decide how to move its body in response to the space around it. Google DeepMind says the model can be adapted to different robot designs, from two-armed platforms to humanoids.
From hands to feet
Gemini Robotics 2 expands control across an entire humanoid body. A robot can adjust its balance as it steps, bends, squats or reaches through cluttered spaces instead of working from a fixed position. The model also supports complex hands and grippers, with demonstrations including delicate actions such as tying knots and screwing in a light bulb.
The aim is adaptability rather than a single repeated routine. Google DeepMind says robots using the model can respond to unfamiliar objects and changing situations, although performance shown in demonstrations does not mean every task or robot platform is ready for unsupervised real-world use.
A reasoning model acts as the high-level brain
Gemini Robotics ER 2 handles higher-level planning. It can interpret everyday commands, understand objects and physical spaces, break a request into multiple steps and track whether a task has succeeded. While it plans, a lower-level action model manages the robot’s movements.
Local operation for unreliable connections
Google DeepMind is also developing Gemini Robotics On-Device 2, a lighter model that runs locally on robotic hardware. Local processing can reduce network delays and allow robots to keep operating when internet access is limited. The company says the model can adapt to a new robot with fewer than 200 examples, though access is currently limited to trusted testers.
Promising technology, but not a consumer robot launch
Gemini Robotics 2 is an intelligence layer for robot developers, not a finished household robot that consumers can buy. The main action model remains in private preview, while Gemini Robotics ER 2 is available in public preview through Google AI Studio and the Gemini API.
Safety will be central to wider deployment. Google DeepMind describes a layered approach covering social behaviour, task-level reasoning and physical safeguards, but also says its research systems have not been tested on every robot. For businesses, the next test will be whether these models can deliver reliable performance around people outside carefully managed demonstrations.
The system can also coordinate several robots. It is designed to recognise the different strengths of available machines, divide a larger job and communicate across a shared workspace. That capability may be useful in warehouses, manufacturing sites and other environments where one robot cannot efficiently complete every part of a workflow.
Sources: Google DeepMind video; Gemini Robotics overview; Gemini Robotics ER 2; robotics safety approach.