Gemini Robotics 2.5

Gemini Robotics 2.5 is a vision-language-action (VLA) model announced on 10 November 2026 by Google DeepMind. In a blog post signed by Carolina Parada, the lab calls it its most capable VLA model to date: it is meant to increase both the range of skills a robot can learn and the speed at which it can acquire them, with improvements the post groups under dexterity, reliability, and safety. DeepMind posted the announcement as @GoogleDeepMind. There is also a model page.

The lab says it evaluated the model on more than 200 distinct tasks — examples include serving food with tongs, tying a bow on a gift box, and pipetting a liquid — and reports a 30% to 100% higher success rate depending on the task type, plus 20% to 30% faster completion, versus previous baselines. As with earlier Gemini Robotics models, it is described as working across robot types, including Apptronik's Apollo 2 humanoid (with the five-finger, 22-DoF SharpaWave hand) and a Franka Duo platform with two-finger parallel grippers.

For a real-world check, DeepMind partnered with Apptronik and Mercedes-Benz on a proof of concept at production lines in Kecskemét, Hungary, and Sindelfingen, Germany. In the lab's account, Apollo 2 used an embodied reasoning model (Gemini Robotics ER) to track the workflow and Gemini Robotics 2.5 to control hands and arms, picking metal brackets from supply bins into sequencing trays, with a 96% success rate. The post says the model is rolling out to DeepMind's robotics partners. Development was accelerated, the lab writes, by Gemini 4 Argon acting as a simulation engine (generating varied practice runs and digital twins from camera scenes).

Highlights:

• Official blog (10 November 2026): https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/ ;
• Official announcement: https://x.com/GoogleDeepMind/status/2120169787896525104 ;
• Model page: https://deepmind.google/models/gemini-robotics/vla/ ;
• Distinct from Gemini 4 Carbon and Gemini 4 Argon, which are general frontier models; this name is the robotics VLA line. The same post also introduces ASIMOV-Jailbreak, an open safety benchmark for robotics jailbreaks (the post cites a 99.5% refusal rate for Argon on that set, versus 82% for a prior Gemini Robotics ER 2 baseline).

Why it matters

Physical AI is moving from lab demos to unstructured workplaces, and a named VLA generation with a factory-floor pilot is a concrete checkpoint on that path. For robotics teams, the useful objects are the official write-up, the model page, and partner access — not a claim that humanoid labor is generally available. The Mercedes-Benz kitting figure is a proof of concept on a specified task, reported by DeepMind and its partners.

The post also stresses steerability: inference-time reasoning guidance is said to have raised safe task completion around obstacles by 33% and around nearby people by 43% in the lab's evaluations. That, plus the new jailbreak benchmark, is the safety story DeepMind is putting next to the capability claims.

Limits

Task counts, success-rate ranges, speed gains, and the 96% kitting figure come from DeepMind (and partner quotes). We have not independently verified them. The model is described as rolling out to partners, not as a generally available API or consumer product. Coverage is of the tasks and sites in the announcement, not of arbitrary factory work or home use. Gemini Robotics ER is a related embodied-reasoning model mentioned in the same post; do not treat it as the same name as Gemini Robotics 2.5.

1. Read the official post https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/ and the announcement https://x.com/GoogleDeepMind/status/2120169787896525104 ;
2. See the model page https://deepmind.google/models/gemini-robotics/vla/ for the product grouping of the Gemini Robotics line ;
3. Keep this distinct from Gemini 4 Carbon / Gemini 4 Argon (general models) and from Gemini Robotics ER (the embodied-reasoning model used alongside 2.5 in the factory pilot) ;
4. Treat the Mercedes-Benz 96% figure as a partner proof of concept on kitting, not a general guarantee, and assume access is through DeepMind's robotics partners unless a public API is later documented.