Gemini Robotics and Gemini Robotics-ER

2025-03-12 · Publisher : Google · Domain : Artificial Intelligence

Google DeepMind introduces Gemini Robotics, a vision-language-action model based on Gemini 2.0 that directly controls robots, and Gemini Robotics-ER, dedicated to spatial reasoning.

Google DeepMind introduces Gemini Robotics, a vision-language-action model based on Gemini 2.0 that directly controls robots, and Gemini Robotics-ER, dedicated to spatial reasoning. ### What changes - Gemini Robotics: a vision-language-action (VLA) model that adds physical action as an output modality to drive a robot. - Gemini Robotics-ER: a Gemini model with advanced spatial understanding, so that roboticists can run their own programs using its embodied reasoning. - According to the technical report cited by Google, Gemini Robotics more than doubles performance on average on a comprehensive generalization benchmark. - Multi-step dexterous tasks (folding origami, packing a snack into a zip bag). - Partnership with Apptronik for humanoid robots; access to Gemini Robotics-ER limited to trusted testers. ### Specifications - Base: Gemini 2.0. Responds to instructions in everyday language, in several languages. - Availability: partners and trusted testers, no public access at launch. ### Pricing Pricing not disclosed at launch. ### Benchmarks No precise numerical score in the post, apart from the mention that average generalization performance "more than doubles" (technical report). ### Key takeaways First version of Gemini for physical robotics. It remains reserved for partners in 2025, which places it out of reach of typical enterprise deployments.

Tags : LLM, Google, DeepMind, Robotique, Gemini