Google DeepMind introduced Gemini Robotics 2.5 on 2026-11-10, describing it as its most capable vision-language-action (VLA) model to date (Deepmind (official statement)). The company stated that the model increases both the breadth of skills a robot can learn and the speed at which it learns them, and that it is rolling out to robotics partners (Deepmind (official statement)). In a proof-of-concept with Apptronik and Mercedes-Benz, Apollo 2 humanoid robots running the model reached a 96% success rate on automotive kitting (Deepmind (official statement)).
Key Facts
| Item | Detail | Source |
|---|---|---|
| Model | Gemini Robotics 2.5, a VLA model | Deepmind (official statement) |
| Task evaluation | Over 200 distinct tasks | Deepmind (official statement) |
| Factory pilot | 96% success rate on Mercedes-Benz lines | Deepmind (official statement) |
| Safety benchmark | ASIMOV-Jailbreak, open source | Deepmind (official statement) |
| Availability | Rolling out to robotics partners | Deepmind (official statement) |
Timeline
- 2026-11-10: Google DeepMind introduced Gemini Robotics 2.5 and released the ASIMOV-Jailbreak benchmark (Deepmind (official statement)).
- 2026-11-10: Ars Technica reported on the release and on the 96 percent factory result (Ars Technica).
Details (as Reported)
DeepMind stated that the model was evaluated on tasks such as serving food with tongs, tying a bow on a gift box and pipetting a liquid (Deepmind (official statement)). Compared with previous baselines, it reported a 30% to 100% higher success rate depending on the task type, with tasks completed 20% to 30% faster (Deepmind (official statement)). Ars Technica noted that these figures are based on laboratory testing (Ars Technica).
The factory pilot ran on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany (Deepmind (official statement)). Apollo 2 sorted parts from bulk bins into trays in the order needed for each vehicle, using Gemini Robotics ER to track the workflow and the Gemini Robotics 2.5 actions model to control its hands and arms (Deepmind (official statement)). The model works with several robot types, including Apollo 2 with the five-finger, 22 degree-of-freedom SharpaWave hand and the Franka Duo platform (Deepmind (official statement)).
On safety, DeepMind reported that inference-time reasoning guidance boosted safe task completion around obstacles by 33% and around nearby people by 43% (Deepmind (official statement)). It also released ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts that disguise unsafe requests, and reported that Gemini 4 Argon reached 99.5% jailbreak resistance against 82% for Gemini Robotics ER 2 (Deepmind (official statement)). The company added that Gemini 4 Argon acted as a simulation engine that accelerated development of the model (Deepmind (official statement)).
What Is Not Yet Confirmed
- Ars Technica reported that a human doing this kind of repetitive task gets it right virtually every time, and that DeepMind director Kanishka Rao called closing the last 4 percent gap a steep road (Ars Technica).
- DeepMind described the Mercedes-Benz result as a pilot showing promising progress toward widespread industrial applications (Deepmind (official statement)).