Google DeepMind releases Gemini Robotics 2.5 for robot dexterity
Google DeepMind introduced Gemini Robotics 2.5 on November 10, 2026, which it calls its most capable vision-language-action (VLA) model to date. A VLA model takes camera input and an instruction and outputs robot actions. DeepMind says 2.5 learns more skills, and learns them faster, than its predecessor, and it is rolling out to robotics partners rather than as a public API.
What’s new
DeepMind says it evaluated the model on more than 200 distinct tasks, among them serving food with tongs, tying a bow on a gift box and pipetting liquid. Against previous baselines it reports success rates 30% to 100% higher, depending on the task type, and tasks finished 20% to 30% faster. It points to tool use, deformable objects and delicate items as the main gains.
DeepMind’s own table compares Gemini Robotics 2 with 2.5 on individual tasks:
| Task | Gemini Robotics 2 | Gemini Robotics 2.5 |
|---|---|---|
| Close the lid of a black toolbox | 0% | 93% |
| Pack pastries with tongs | 5% | 100% |
| Untie ribbon bows | 0% | 70% |
| Unhang a mug to a table | 29% | 75% |
| Hang a cardigan | 0% | 62% |
| Sweep trash into a dustpan with a brush | 0% | 52% |
| Pack a rice cooker | 0% | 40% |
| Fold pants | 0% | 40% |
These are DeepMind’s figures from its own evaluations, not independent tests. Several tasks that Gemini Robotics 2 could not do at all still succeed less than half the time with 2.5.
The model works natively with more than one robot body. DeepMind names the Apptronik Apollo 2 humanoid, with a five-finger, 22-degree-of-freedom SharpaWave hand, and the Franka Duo platform, which uses two-finger parallel grippers.
A car factory test
DeepMind tested the model with Apptronik and Mercedes-Benz on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 did kitting: picking parts out of bulk bins and sorting them into ordered trays for each vehicle. DeepMind reports a 96% success rate. Gemini Robotics ER tracks the workflow while Gemini Robotics 2.5 controls the hands and arms. Apptronik CEO Jeff Cardenas said, “we saw the models improve dramatically in just a few weeks of post-training.”
DeepMind describes this as a proof of concept, so it says nothing yet about production deployment or scale.
How Gemini 4 Argon was used
DeepMind says Gemini 4 Argon sped up development by acting as a simulation engine. From a real camera scene it builds a digital twin and generates thousands of varied practice runs. Combined with real demonstrations, that helps the robot generalize to new scenes and objects. Researchers also use it to diagnose failed attempts from video and test fixes in simulation and on real robots.
Safety numbers
The model uses inference-time reasoning guidance to avoid physical hazards. In DeepMind’s evaluations this raised safe task completion by 33% around obstacles and 43% around nearby people.
DeepMind also released ASIMOV-Jailbreak, an open-source benchmark on Hugging Face with hundreds of prompts that disguise unsafe requests across dozens of scenes. It reports 99.5% jailbreak resistance for Gemini 4 Argon, against 82% for the earlier baseline, Gemini Robotics ER 2.
What it means for developers
There is no public API, pricing or general release date in the announcement. Access is through robotics partners; DeepMind names Apptronik, Mercedes-Benz and Agile. If you build robot software, the practical items today are the open ASIMOV-Jailbreak benchmark, which you can use to test your own embodied models, and the split DeepMind describes: a planning model (Gemini Robotics ER) that tracks a workflow and an action model (Gemini Robotics 2.5) that drives the hardware.
Google has not said when, or whether, developers will get direct access. Until it does, treat the success rates above as vendor-reported results on DeepMind’s chosen tasks.
For another recent Google AI infrastructure announcement, see our post on Google’s TPU 8tx and TPU 8i Max.