Gemini Robotics 2.5
Google DeepMind released Gemini Robotics 2.5, a vision-language-action model with a 30% to 100% higher task success rate than its baselines, and the open ASIMOV-Jailbreak safety benchmark.
Gemini Robotics 2.5 was evaluated on over 200 tasks and completes them 20% to 30% faster than previous baselines. It adds inference-time reasoning guidance that raised safe task completion by 33% around obstacles and 43% around nearby people. DeepMind says Gemini 4 Argon sped up development by generating simulated practice runs and digital twins from real camera scenes.
- Date
- Tuesday 10 November 2026
- Lab
- Google DeepMind
- Kind
- model
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Success rate gain vs previous baselines | 30% to 100% depends on task type, over 200 tasks | company |
| Task completion speed gain | 20% to 30% faster vs previous baselines | company |
| Apollo 2 automotive kitting success rate at Mercedes-Benz | 96% proof-of-concept on production lines | company |
| Safe task completion gain around obstacles | 33% inference-time reasoning guidance | company |
| Safe task completion gain around nearby people | 43% inference-time reasoning guidance | company |
| Jailbreak resistance with Gemini 4 Argon on ASIMOV-Jailbreak | 99.5% vs 82% for Gemini Robotics ER 2 per DeepMind blog | company |
Not generally available. It is rolling out to robotics partners such as Apptronik, Agile Robots and Boston Dynamics. The ASIMOV-Jailbreak benchmark is open source on Hugging Face.
Sources
- deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/
- arstechnica.com/google/2026/11/deepmind-reveals-gemini-robotics-2-5-and-open-safety-benchm
- mlq.ai/news/google-deepmind-releases-gemini-robotics-2-5-to-partners-and-an-open-asimov-ja
- siliconangle.com/2026/11/11/google-deepmind-releases-new-advanced-robotics-ai-model-and-ne
This record was checked against its sources on 11 November 2026. How we check