Google DeepMind has released Gemini Robotics 2.5, which it calls its most capable vision-language-action model so far. In a blog post dated 10 November, robotics lead Carolina Parada said the model is meant to widen the range of skills a robot can learn, and to help it acquire them faster, as the company moves robots “out of the lab and into more unstructured, real-world environments.” DeepMind says the model is rolling out to its robotics partners.
What DeepMind is claiming
Gemini Robotics 2.5 is the action model in DeepMind’s robotics stack: it turns camera input and language into motor commands. A separate embodied-reasoning model, Gemini Robotics ER, is used for planning and keeping track of a workflow. DeepMind’s product page still lists Gemini Robotics ER 2 and a lighter On-Device 2 model alongside the new 2.5 release. This is a follow-up to Gemini Robotics 2, announced in July, not a new Gemini chatbot.
DeepMind says it evaluated the model across more than 200 distinct tasks, including serving food with tongs, tying a bow on a gift box, and pipetting a liquid. Compared with “previous baselines,” it reports a 30% to 100% higher success rate depending on the type of task, and says robots completed tasks 20% to 30% faster. Those figures are DeepMind’s own results; the post does not name the exact prior model used as the baseline for every test, and it does not present an independent benchmark. The company also says the model can run on different robot bodies, including Apptronik’s Apollo 2 humanoid, with a five-finger, 22-degree-of-freedom SharpaWave hand, and a Franka Duo setup with two-finger parallel grippers.
| Claim | What DeepMind reported |
|---|---|
| Task breadth | Evaluated across more than 200 distinct tasks |
| Success rate | 30% to 100% higher than previous baselines, depending on the task |
| Speed | 20% to 30% faster task completion |
| Factory trial | 96% success rate on automotive kitting with Apollo 2 |
| Safety around obstacles | 33% higher safe task completion |
| Safety around people | 43% higher safe task completion |
A factory trial with Apptronik and Mercedes-Benz
The most concrete test DeepMind describes is a proof of concept on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apptronik’s Apollo 2 robots, running Gemini Robotics ER to track the workflow and Gemini Robotics 2.5 to control hands and arms, picked metal automotive brackets from supply bins and placed them into sequencing trays in the order needed for each vehicle, a task known as kitting. DeepMind says the robots reached a 96% success rate and could correct for gaps in the trays.
Jeff Cardenas, co-founder and chief executive of Apptronik, is quoted in DeepMind’s post as saying the models improved “dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off.” Mercedes-Benz has not issued a same-day statement of its own. In an interview with Ars Technica, DeepMind director Kanishka Rao said closing the remaining 4% gap would be a “steep road.” Parada told the publication that the team could “be very good at a particular application at the cost of generalization,” or “push on the generalization axis and maybe not get to 99.9 in every application.” Ars also reported that the factory video DeepMind shared is sped up 2×, and that the current 2.5 model still runs on cloud servers rather than on the robot itself.
Gemini 4 Argon as a training aid
DeepMind says Gemini 4 Argon helped it get around a shortage of physical training data by acting as a “simulation engine.” From a real-world camera scene, Argon can generate thousands of varied simulated practice runs and a digital twin, which are then combined with real demonstrations. The company also says Argon can suggest why a task failed from video, and that those ideas can be tested in simulation and on real robots. It describes the 2.5 release as a milestone “on the path toward solving physical AGI.”
Safety steerability and ASIMOV-Jailbreak
On safety, DeepMind says Gemini Robotics 2.5 uses “inference-time reasoning guidance” to steer around physical hazards as it runs. In its evaluations, that raised safe task completion around obstacles by 33% and around nearby people by 43%.
The company is also releasing ASIMOV-Jailbreak, a new open-source benchmark of “hundreds of prompts that disguise unsafe requests across dozens of scenes.” DeepMind says Gemini 4 Argon currently leads its agentic robotics evaluations on unsafe-task refusal, resolving uncertainty, and working near people, and that Argon reaches 99.5% jailbreak resistance on the new test, compared with 82% for Gemini Robotics ER 2. Those scores are for Argon and the earlier embodied-reasoning model, not a published score for Gemini Robotics 2.5 itself. The blog post links a Hugging Face dataset, google/asimov_agentic, though access to that repository is restricted.
Who can use it
Gemini Robotics 2.5 is not a general public release. DeepMind says it is rolling the model out to robotics partners, and the product page invites interested teams to join a waitlist. The blog post thanks Agile, Apptronik and Mercedes-Benz. Some onlookers on X had been expecting a Gemini 4 Carbon announcement at the same time; DeepMind’s post was about the robotics model.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
