Google ships Gemini Robotics 2.5, a VLA model for real-world dexterity
DeepMind released Gemini Robotics 2.5 on November 10, 2026 and calls it its most capable vision-language-action model. A VLA model turns what a robot's cameras see directly into movement, with no separate planning stage between perception and the arm. Google says 2.5 widens both the range of tasks a robot can learn and the speed at which it learns them. The company tested the model on more than 200 distinct tasks and is rolling it out to robotics partners, including Apptronik, Agile, and Mercedes-Benz.
200 tasks, from tongs to pipettes
The evaluation covers 200 distinct tasks, with examples like serving food with tongs, tying a bow on a gift box, and pipetting a liquid. Google chose them to combine three skills at once: using tools, manipulating deformable objects, and handling delicate items with finesse.
By the company's own numbers, 2.5 posts a 30% to 100% higher success rate than previous baselines, depending on the task type, and completes the same tasks 20% to 30% faster. Google also says the testing approach could scale to many more tasks than the 200 it ran.
Like the earlier Gemini Robotics models, 2.5 works natively across several robot platforms: Apptronik's Apollo 2 humanoid, which carries the five-finger, 22-degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-fingered parallel grippers.
96% on Mercedes-Benz lines
The most concrete data point in the post is an industrial proof-of-concept. Google partnered with Apptronik and Mercedes-Benz to run Apollo 2 humanoids on active automotive production lines at two plants, Kecskemét in Hungary and Sindelfingen in Germany.
The task is automotive kitting: the robot pulls metal brackets out of bulk bins and places them into sequencing trays in the exact order each vehicle on the line needs. The Apollo 2 pairs two models: Gemini Robotics ER, Google's embodied reasoning model, keeps track of the workflow, while Gemini Robotics 2.5 controls the hands and arms, correcting for gaps as it works.
The pilot hit a 96% success rate, according to Google. Jeff Cardenas, Apptronik's co-founder and CEO, says the models improved dramatically in a few weeks of post-training and carried over to Mercedes-Benz production with very little drop-off.
Safety: steerability and a jailbreak benchmark
The release also pushes on physical safety. 2.5 uses inference-time reasoning guidance to steer around physical hazards in real time, a layer Google describes as analogous to the safety classifiers that wrap Gemini's responses, applied to robot motion. In the company's evaluations, that lifted safe task completion around obstacles by 33% and by 43% when a person is nearby.
Google also released ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts that disguise unsafe requests for robots across dozens of scenes. Against it, Gemini 4 Argon posts 99.5% jailbreak resistance, up from the 82% of the previous Gemini Robotics ER 2 baseline. All of these numbers come from Google's own post and have not been independently reproduced as of the release date.
Argon as a research accelerator
The post devotes a section to how Gemini 4 Argon sped up 2.5's development. Argon acts as a simulation engine: from a real camera scene it generates thousands of varied practice runs and builds a realistic digital twin. Combined with actual demonstrations, that lets the robot generalize to new scenes and objects faster. Argon also reviews footage of failed tasks, hypothesizes why they failed, and lets researchers test fixes in simulation and on the real robot.
The line's cadence is fast. The first Gemini Robotics shipped in March 2025, 1.5 in September 2025, and 2, which added whole-body intelligence for humanoids, in July 2026. Google calls 2.5 a milestone on the path to physical AGI. That is the company's framing, not a measurable claim.
What to verify
2.5 is rolling out to robotics partners. If you are evaluating a VLA model for a factory or a lab, the kitting pilot is the number to scrutinize first: 96%, one task, two plants, Google's own figures. The ASIMOV-Jailbreak benchmark is the part you can check yourself, since it is open source. The raw data behind the 200-task evaluation has not been published as of the release date.
Source: deepmind.google
More news
Have something in mind?
Start a conversation