Gemini Robotics

Gemini Robotics is a family of artificial intelligence models from Google DeepMind designed to control robots. The family includes a vision-language-action (VLA) model that converts images and natural-language instructions directly into motor commands, and an embodied reasoning model, Gemini Robotics-ER, that plans and reasons about space, objects and tasks before any motion happens.^[1]

History

Google DeepMind introduced Gemini Robotics and Gemini Robotics-ER on 11 March 2025, both built on Gemini 2.0. The VLA model added physical actions as a new output modality alongside text, images and video, while the ER model extended Gemini's spatial understanding — pointing, 3D detection, trajectory and grasp prediction — for roboticists to plug into their own low-level controllers.^[1][2] In 2025 DeepMind released the next generation, Gemini Robotics 1.5 and Gemini Robotics-ER 1.5, introducing a Thinking VLA that reasons in natural language before acting and a Motion Transfer mechanism that lets one model learn across different robot bodies. Gemini Robotics-ER 1.5 became available to developers through the Gemini API in Google AI Studio, with the VLA offered to select partners.^[3][4]

Key Concepts and Technology

A vision-language-action model takes in camera frames and a written instruction and outputs an action chunk — a short sequence of joint movements — rather than generating code or text. Gemini Robotics performs this at roughly 50 Hz end-to-end: a cloud-hosted backbone, a distilled version of the ER model, responds in under 160 ms, while an onboard action decoder running on the robot smooths the latency into reactive motion, for about 250 ms from observation to action.^[2]

Embodied reasoning is the second pillar. Gemini Robotics-ER handles perception, state estimation, spatial understanding, planning and code generation in one model — for example, seeing a coffee mug and inferring a two-finger grasp on the handle plus a safe approach trajectory. In an end-to-end setting this achieved roughly 2–3× the task success rate of plain Gemini 2.0.^[1] DeepMind released the ERQA benchmark (Embodied Reasoning Question Answering), 400 multiple-choice visual questions covering spatial, trajectory, action and state-estimation reasoning, to measure such capabilities.^[2]

Version 1.5 added embodied thinking — the model writes out its reasoning in natural language before emitting an action, decomposing complex instructions into primitive skills and making its behaviour more interpretable — and motion transfer, which allows skills learned on one robot, such as the ALOHA 2 research platform, to run on a different body like Apptronik's Apollo humanoid without robot-specific retraining.^[3][4]

Applications and Impact

Demonstrated capabilities include folding an origami fox, packing lunch bags, playing cards and picking up unseen objects in cluttered settings; new short-horizon skills can be learned from as few as 100 human demonstrations.^[2] The models target domestic assistance, warehouse manipulation, electronics assembly and laboratory automation. DeepMind has paired the releases with research into safety and alignment for agentic robots operating around people, and the ER model's ability to generate robot code on the fly lowers the barrier for labs that previously had to compose perception, planning and control stacks from separate systems.^[1][4]

>See Also

• Gemini
• Physical AI
• Humanoid Robots
• Vision-Language-Action Model
• Google DeepMind