Many MindedDeepMind's new robot model handles 200+ tasks and a car-factory pilot
Google DeepMind announced Gemini Robotics 2.5 on November 10, 2026, calling it its most capable vision-language-action model yet. It was tested on more than 200 distinct tasks, from serving food with tongs to tying a bow on a gift box to pipetting liquid. DeepMind reports success rates 30% to 100% higher than previous baselines, depending on task type, and tasks finished 20% to 30% faster. The model is rolling out to DeepMind's robotics partners.
The model runs across different robot bodies, including Apptronik's Apollo 2 humanoid, which has a five-finger, 22-degree-of-freedom SharpaWave hand, and the two-gripper Franka Duo platform. In a proof-of-concept with Apptronik and Mercedes-Benz, Apollo 2 robots worked on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. They did automotive kitting, moving parts such as metal brackets from bulk bins into sequenced containers, with a 96% success rate. Gemini Robotics ER tracked the workflow while the 2.5 actions model controlled hands and arms.
Safety work was part of the release. Reasoning guidance at inference time raised safe task completion by 33% around obstacles and 43% around nearby people. DeepMind also open-sourced ASIMOV-Jailbreak, a benchmark of hundreds of prompts that hide unsafe requests across dozens of scenes. Gemini 4 Argon resisted 99.5% of them, versus 82% for the earlier Gemini Robotics ER 2.
Why it matters: the hard part of general-purpose robots is dexterity, and one model that transfers across hands and bodies could spread the cost of that skill over many machines. If it holds up, physical labor in manufacturing, and eventually care work, gets cheaper.
Caveats: every figure comes from DeepMind's own evaluations and has not been independently verified. The gains are measured against unspecified baselines. The 200 tasks are a test set, and the Mercedes-Benz work is a pilot, not a full production rollout.
Source: Google DeepMind