Google DeepMind introduces Gemini Robotics and embodied-reasoning models

Google DeepMind introduced Gemini Robotics and Gemini Robotics-ER on 12 March 2025, extending Gemini 2.0 into robot control and embodied reasoning. Gemini Robotics converts visual observations and language instructions into physical actions. Robotics-ER supplies spatial reasoning and planning that developers can connect to their own robot controllers. Access initially involves selected collaborators and trusted testers.

The vision-language-action model was trained primarily on ALOHA 2 dual-arm data, with demonstrations also extending to a Franka-based system. Tasks included folding paper and placing a snack inside a bag. DeepMind showed the robot adapting when objects moved or instructions changed, illustrating how visual feedback can update an action during execution.

DeepMind also announced a partnership with Apptronik to develop humanoids using Gemini 2.0, including adaptation to Apollo. Robotics-ER is intended to support perception, state estimation, planning and code generation alongside existing low-level control systems. The launch included the ASIMOV dataset for evaluating semantic safety in robotic actions.