Google DeepMind on November 10, 2026 announced Gemini Robotics 2.5, which it calls its most capable vision-language-action (VLA) model so far. The model controls a robot's arms and hands directly. DeepMind says it improves dexterity, reliability and safety, and lets robots learn new skills faster.
DeepMind says it evaluated the model on 200 distinct tasks. Depending on the type of task, success rates were 30% to 100% higher than previous baselines, and tasks finished 20% to 30% faster. Some individual tasks improved sharply: packing pastries with tongs went from 5% to 100% success, untying ribbon bows from 0% to 70%, hanging a cardigan from 0% to 62%, and folding pants from 0% to 40%.
In a pilot with Mercedes-Benz, an Apptronik Apollo 2 humanoid running the models did automotive kitting at a 96% success rate. In that setup the Gemini Robotics ER embodied-reasoning model tracked the workflow and Gemini Robotics 2.5 controlled the robot's hands and arms. DeepMind says navigating around obstacles improved safe task completion by 33%, and navigating around nearby people by 43%. It also reports 99.5% jailbreak resistance for Gemini 4 Argon, against 82% for Gemini Robotics ER 2, and has open-sourced the ASIMOV-Jailbreak benchmark on Hugging Face.
The model runs on the Apollo 2, which uses a five-finger SharpaWave hand with 22 degrees of freedom, and on the Franka Duo platform with two-fingered grippers. DeepMind says it works natively with many robot types.
Gemini Robotics 2.5 is rolling out to robotics partners, and DeepMind names Agile, Apptronik and Mercedes-Benz. The announcement does not mention public API access. For now, developers outside these partnerships can't use the model directly. The results show a single model handling fine manipulation on both humanoid hands and simpler grippers.