Google DeepMind announced Gemini Robotics 2.5 on 10 November 2026, describing it as its most capable vision-language-action (VLA) model to date. It is rolling out to robotics partners rather than as a general release.
What happened
In a post dated 10 November 2026, Google DeepMind introduced Gemini Robotics 2.5, a model that turns visual and language input into robot motor control. DeepMind says it aims to widen the range of skills robots can learn and speed up how quickly they learn them. The release builds on the Gemini Robotics 2 family; the earlier generation and its safety benchmark are covered in our earlier report on Gemini Robotics 2.
Key details
- Dexterity. DeepMind says it tested more than 200 distinct tasks, including tool use, deformable objects and delicate handling. Against previous baselines it reports success rates 30% to 100% higher depending on task type, and task completion 20% to 30% faster.
- Industrial pilot. In a proof of concept on Mercedes-Benz production lines at Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik’s Apollo 2 humanoid reached a 96% success rate picking and packing automotive parts. In that setup the Gemini Robotics ER embodied-reasoning model tracks the workflow while Gemini Robotics 2.5 controls the hands and arms. This is a vendor-reported pilot result, not an independent evaluation.
- Safety. DeepMind reports that inference-time reasoning guidance raised safe task completion by 33% around obstacles and 43% around nearby people in its evaluations.
- ASIMOV-Jailbreak. DeepMind released an open-source benchmark of hundreds of prompts that disguise unsafe requests. It reports that its agentic model Gemini 4 Argon reached 99.5% jailbreak resistance, against 82% for the previous baseline, Gemini Robotics ER 2.
- Training method. DeepMind says Gemini 4 Argon acted as a simulation engine, generating varied simulated practice from real camera scenes, building digital twins and helping diagnose task failures from video.
- Hardware. The model is described as working across robot types, including Apollo 2 (with a five-finger, 22-degree-of-freedom SharpaWave hand) and the Franka Duo platform with two-finger grippers.
- Access. DeepMind’s announcement says Gemini Robotics 2.5 is rolling out to robotics partners and names Apptronik, Mercedes-Benz and Agile. The announcement does not describe a public API. On DeepMind’s Gemini Robotics page, the separate Gemini Robotics ER 2 model is available as a preview in Google AI Studio (gemini-robotics-er-2-preview), and the page cites 100+ trusted testers and a waitlist for early access to the models.
Media coverage reported by outlets such as Ars Technica, as summarised by aggregators, notes the model still depends on cloud inference, which adds latency, and that reasoning-based hazard detection adds overhead. We have not independently verified those characterisations.
What it means for researchers and research offices
- Access is the constraint. With rollout limited to partners, most academic robotics labs cannot yet run Gemini Robotics 2.5 directly. The ER 2 preview in AI Studio is the more accessible entry point for embodied-reasoning and planning work.
- Benchmarks are comparable only on DeepMind’s terms. The headline figures (96%, 99.5%, 30% to 100%) are self-reported against DeepMind-chosen baselines. Labs citing them should attribute them and avoid treating them as independently replicated.
- ASIMOV-Jailbreak is reusable. Because it is open source, safety and robotics groups can evaluate their own systems against it, which may help with the safety documentation that institutional review and funders increasingly ask for.
- Partnership terms matter. Research offices negotiating industry collaborations or trusted-tester access should review data-use, publication and IP terms, since partner-only model access often comes with restrictions on what can be published.
What to watch
- Whether DeepMind widens access beyond named partners, for example through trusted-tester or API availability for 2.5 itself.
- Independent evaluations of the 96% pilot figure and of the jailbreak-resistance claims.
- Adoption of ASIMOV-Jailbreak by groups outside DeepMind.
Sources
- Google DeepMind, Gemini Robotics 2.5 accelerates real-world physical intelligence (10 Nov 2026), read 13 Nov 2026
- Google DeepMind, Gemini Robotics models page, read 13 Nov 2026
- Automate.org, Google DeepMind releases Gemini Robotics 2.5, read 13 Nov 2026
- Let’s Data Science, DeepMind releases Gemini Robotics 2.5 and open ASIMOV-Jailbreak safety benchmark, read 13 Nov 2026
