Google DeepMind has introduced Gemini Robotics 2.5, connecting visual information and spoken instructions to robot movements. Trials with Mercedes-Benz involve selecting metal components and arranging them for assembly in Hungary and Germany. The work combines different robot hands, physical demonstrations and simulation, with access through partners rather than a general release for every robot platform.
Google DeepMind introduced Gemini Robotics 2.5 as a new model connecting what a robot sees and the instructions it receives to physical movements. Partner work includes different hand and gripper designs, from two-arm systems to humanoid robots. Apptronik, a US humanoid-robot developer, is working with the model through its Apollo 2 robot. The release is being introduced through robotics partners.[1], [2]
At Mercedes-Benz facilities in Kecskemét, Hungary, and Sindelfingen, Germany, pilots involve selecting metal brackets from mixed bins and putting them into sequenced trays for assembly. Gemini Robotics-ER, DeepMind’s robot-reasoning model, follows the workflow and handles higher-level reasoning. The action model directs the robot’s hands and arms. This division connects an ordered factory task to the individual movements needed to complete it.[1]
Training draws on physical demonstrations, simulated environments and digital twins. DeepMind describes varying scenes and camera views to broaden the situations robots encounter, and testing explanations for failures in both simulated and physical settings. Its ASIMOV-Jailbreak evaluation resource examines unsafe requests concealed inside apparently harmless instructions. That language-level safety work accompanies partner pilots; it does not establish independently verified physical safety across all environments.[1]
