DeepMind unveils Gemini Robotics 2.5 at CoRL with 30–100% higher task success

Google DeepMind launched Gemini Robotics 2.5 alongside CoRL in Austin and delivered it to enterprise partners. The release improves dexterity and real-world task completion. Partners report 30% to 100% higher task success rates in physical robotics evaluations, and Google says its robotic AI is getting close to human-level completion of real-world tasks.

The proof point Google highlights is deployment. Gemini-powered robots have already carried out work at a Mercedes-Benz factory, moving beyond lab demonstrations. The model family pairs Gemini's multimodal reasoning with action generation, so robots can interpret instructions and scenes and then execute manipulation steps.

Robotics foundation models are a crowded race, with competitors including Physical Intelligence, Figure, Tesla's Optimus effort and NVIDIA's robotics stack. Google's distinctive bet is reusing frontier Gemini reasoning across chat, code and robots. Releasing it on the same day as Gemini 4 Carbon reinforces that platform story.

The caveats are familiar. The 30–100% range depends heavily on which tasks partners measured, and "close to human-level" is Google's own characterization without a standard public benchmark. Watch for disclosed factory throughput numbers and whether more manufacturers sign on.