Google DeepMind today introduced Gemini Robotics, a new AI model designed to control robots directly, and Gemini Robotics-ER, a companion model for embodied reasoning. Both are built on Gemini 2.0, extending the model’s multimodal capabilities from the digital realm into the physical world.

Gemini Robotics is a vision-language-action (VLA) model that outputs physical actions to control robot arms and hands. According to DeepMind, the model more than doubles performance on a generalization benchmark compared to other state-of-the-art VLA models. It demonstrates the ability to perform dexterous tasks such as origami folding and packing snacks into Ziploc bags, and it can adapt to changes in its environment, like replanning when an object slips from its grasp.

The model operates across multiple robot types, including the bi-arm ALOHA 2 platform and, through a partnership with Apptronik, the humanoid Apollo robot. Separately, Gemini Robotics-ER improves spatial understanding and can generate code for robot control on the fly, achieving a 2x-3x success rate over Gemini 2.0 in end-to-end settings.

DeepMind emphasized a layered safety approach, including a new ASIMOV dataset for evaluating semantic safety in robotics. The company also revived a concept inspired by Isaac Asimov’s Three Laws, using a ‘Robot Constitution’ framework to generate rules that steer robot behavior.

On Hacker News, the announcement drew 933 points and 561 comments. Much of the discussion centered on how the model evokes Asimov’s fictional laws, with another commenter calling Asimov “The pioneer of AI alignment.” while others debated the practicality of general-purpose robots and the risks of combining large language models with physical hardware.