Gemini Robotics 2.5: DeepMind boosts robot dexterity and reliability in real-world settings

Google DeepMind unveiled Gemini Robotics 2.5, which it describes as its most capable vision-language-action (VLA) model to date. According to the company, the model broadens both the range of skills a robot can learn and how quickly it acquires them. The stated goal is to get robots out of the lab and working in less structured environments, where they can help people with a wider range of tasks. The model is being rolled out to DeepMind's robotics partners. The blog post is dated 10 November 2026.

DeepMind frames this launch as the next step in its series of robotics models. The related posts section lists Gemini Robotics (March 2025), Gemini Robotics 1.5 (September 2025), Gemini Robotics 2 (July 2026), which according to the blog brought full-body intelligence to many types of robots, including humanoids, and Gemini 4 Argon for Robotics (October 2026). The company credits much of the current progress to Gemini 4 Argon's ability to accelerate robotics research.

The post groups the improvements along three axes: dexterity, reliability and safety. On dexterity, DeepMind says its robots can perform significantly more tasks with better performance, from using everyday tools to handling delicate or flexible objects. The evaluation covered more than 200 distinct tasks, and the company notes the approach could scale to many more. Examples cited include serving food with tongs, tying a bow on a gift box and pipetting a liquid. These tasks combine advanced tool use, complex manipulation of deformable objects and fine handling of delicate items, and DeepMind presents them as a step toward human-level dexterity.

On performance, the company says that, compared with its previous baselines, Gemini Robotics 2.5 achieves a success rate between 30% and 100% higher depending on the task type, and completes tasks between 20% and 30% faster. Like earlier models in the family, it works natively across different robot types. The blog mentions Apptronik's Apollo 2 humanoid, fitted with the five-fingered SharpaWave hand with 22 degrees of freedom, and the Franka Duo platform with two-fingered parallel grippers.

The most striking reliability test is a proof of concept with Apptronik and Mercedes-Benz. Apollo 2 robots were used on live production lines in Kecskemét (Hungary) and Sindelfingen (Germany). The robot autonomously performs automotive «kitting». This task involves sorting loose parts from bulk containers into organized trays, in the exact order each vehicle on the line requires. DeepMind describes it as repetitive and physically demanding. According to the company, offloading this preparation work helps keep the assembly line running and lets technicians focus on more complex build tasks.

In the technical setup, the Gemini Robotics ER embodied reasoning model tracks the workflow, while the Gemini Robotics 2.5 action model controls the hands and arms. The robot reaches into supply containers, picks up metal brackets and places them in sequencing containers, filling any gaps it detects. DeepMind reports a 96% success rate and presents it as promising progress toward large-scale real-world industrial applications. It should be noted that this is a proof of concept or pilot, not a production deployment, and the blog does not detail the methodology behind that figure.

Jeff Cardenas, co-founder and CEO of Apptronik, says in the post that the models improved markedly after just a few weeks of post-training and that performance carried over to Mercedes-Benz production with very little loss. He sees this as the foundation for bringing humanoids into industrial and commercial settings at scale.

On safety, DeepMind says it has significantly advanced the model's physical steerability. Gemini Robotics 2.5 uses inference-time reasoning guidance to navigate physical hazards in real time. The company likens this to the safety classifiers that add a protective layer to Gemini's responses. In its evaluations, this capability improved safe task completion by 33% around obstacles and by 43% near people.

For long-horizon agentic tasks, the blog says Gemini 4 Argon currently leads DeepMind's agentic robotics evaluations in refusing unsafe tasks, resolving uncertainty and working in close proximity to humans. The company is also releasing ASIMOV-Jailbreak, a new open-source benchmark with hundreds of instructions that disguise unsafe requests across dozens of scenes, to measure robustness against robotics-specific jailbreaks. According to DeepMind, Argon achieves 99.5% resistance, compared with 82% for its previous baseline, Gemini Robotics ER 2. These are results from the company's own evaluations.

Another central point is Gemini 4 Argon's role in the development itself. DeepMind explains that teaching robots new skills requires large volumes of physical data and efficient modeling algorithms. According to the company, Argon's robotics research capabilities helped resolve those bottlenecks by acting as a simulation engine. Argon can generate thousands of varied simulated rollouts from a real scene captured on camera and build a realistic digital twin. Combined with real training demonstrations, this lets the robot generalize faster to new scenes and objects. Argon can also help researchers form hypotheses about why tasks fail based on video, and test improvements both in simulation and directly on real robots.

DeepMind closes by presenting Gemini Robotics 2.5 as a milestone on the path toward «physical AGI», with advances in dexterity, real-world reliability and rigorous safety. All figures come from the company itself and its partners, and the text includes no independent evaluation or details of the baselines compared. Partners acknowledged include Agile, Apptronik and Mercedes-Benz.