On 10 November 2026, Google DeepMind introduced Gemini Robotics 2.5 on its official blog. The blog describes it as its most capable vision-language-action (VLA) model to date; the Gemini Robotics 2.5 product page calls it "Our most advanced VLA model." The official blog says, "Gemini Robotics 2.5 is rolling out to our robotics partners." The product page's call to action is "Join waitlist for early access." This is not a general public or all-developer general availability release today, and it is not a new Gemini text API model, a price-list update, or a new listing on Vertex or OpenRouter.
The blog says the model increases both the breadth of skills a robot can learn and the speed at which it can acquire them, and it highlights three areas: dexterity, reliability, and safety. Google DeepMind writes that it tested Gemini Robotics 2.5 across over 200 distinct tasks in task breadth, performance, and complexity — examples include serving food with tongs, tying a bow on a gift box, and pipetting a liquid — and that this approach could scale to many more.
Compared with previous baselines, the official blog says Gemini Robotics 2.5 achieves a 30% to 100% higher success rate depending on the type of task, and completes tasks 20% to 30% faster. That is a summary sentence in the official blog, not the same set of figures as the per-task success rates labeled on the official bar charts below.
The official blog publishes bar charts comparing Gemini Robotics 2.5 with predecessor Gemini Robotics 2 on DeepMind's own task suites. These are in-house success-rate results, not a third-party public leaderboard. The light-colored charts from the official blog are shown here.
Official Google DeepMind chart for tabletop pick-and-place: success rates for Gemini Robotics 2.5 versus Gemini Robotics 2 on DeepMind's own task suite, not a third-party public leaderboard.
Official Google DeepMind chart titled "Multi-step & contact-rich," comparing success rates for Gemini Robotics 2.5 versus Gemini Robotics 2 on DeepMind's own task suite, not a third-party public leaderboard. Labels on the chart include Close lid of black toolbox (0% to 93%) and Unhang mug to table (29% to 75%). Those are per-task figures on the chart, not the 30% to 100% or 20% to 30% summary above.
Official Google DeepMind chart for precision kitting and insertion: success rates for Gemini Robotics 2.5 versus Gemini Robotics 2 on DeepMind's own task suite, not a third-party public leaderboard.
Official Google DeepMind chart for dexterous manipulation, comparing Gemini Robotics 2.5 with Gemini Robotics 2 on DeepMind's own task suite, not a third-party public leaderboard. Labels on the chart include packing pastries with tongs (5% to 100%), untying ribbon bows (0% to 70%), hanging a cardigan (0% to 62%), and folding pants (0% to 40%). Those are per-task figures on the chart, not the 30% to 100% or 20% to 30% summary above.
The blog also says that, like previous Gemini Robotics models, Gemini Robotics 2.5 natively works with many different robot types, including Apptronik's Apollo 2 humanoid robot (featuring the five-finger, 22 degree-of-freedom SharpaWave hand) and the Franka Duo platform with two-fingered parallel grippers.
On reliability, the blog describes a real-world proof-of-concept and pilot with Apptronik and Mercedes-Benz, using Apollo 2 humanoid robots on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. In that proof-of-concept, Apollo 2 autonomously performs automotive kitting: sorting individual parts from bulk bins into organized trays in the order needed for each vehicle on the line. The official blog reports a 96% success rate. This is a proof-of-concept / pilot, not general availability or a full production rollout. The blog says the pilot uses the embodied reasoning model (Gemini Robotics ER) to keep track of the workflow, and the Gemini Robotics 2.5 actions model to control the robot's hands and arms.
This official photograph from the Google DeepMind blog shows humanoid robots on an automotive factory assembly line. It is a factory photograph, not a benchmark chart.
For safety, the blog says Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards in real time. In Google DeepMind's evaluations, this capability boosted safe task completion around obstacles by 33% and around nearby people by 43%. The blog also releases ASIMOV-Jailbreak, an open-source benchmark containing hundreds of prompts that disguise unsafe requests across dozens of scenes, with a companion dataset, google/asimov_agentic, on Hugging Face. The 99.5% versus 82% jailbreak-resistance comparison in the blog is for Gemini 4 Argon against Gemini Robotics ER 2; it is not a Gemini Robotics 2.5 score. The blog says Gemini 4 Argon's robotics research capabilities, used as a simulation engine, helped accelerate development of Gemini Robotics 2.5.
The Gemini Robotics product-line hub — not the VLA product subpage — says Google DeepMind is working with 100+ trusted testers, and it lists research partners Agile Robots, Apptronik, and Boston Dynamics. That wording describes the Gemini Robotics product line. It does not mean Gemini Robotics 2.5 is already open to all testers or to the public.
This site already has an introduction to Gemini Robotics ER 2, the embodied reasoning model that previewed on Google Cloud and the Gemini API in late October 2026. The official blog places the ER model and the 2.5 actions model side by side in the Mercedes-Benz pilot, with different roles; that is not the same as saying ER 2 was iterated into 2.5. See the existing Gemini Robotics ER 2 introduction (not today's release; an embodied reasoning model, not the same model as today's VLA 2.5).