Google DeepMind introduced Gemini Robotics 2.5 on 10 November 2026, reporting gains in dexterity across more than 200 tasks and a 96% success rate in a pilot on Mercedes-Benz production lines in Hungary and Germany. The release updates DeepMind’s robotics models, which it says can control different kinds of machines, and gives its robotics partners a model intended to turn what a robot sees and is asked to do into physical actions.
Key points
- In DeepMind’s evaluation across more than 200 tasks, success rates were 30% to 100% higher than previous baselines, depending on the task.
- DeepMind says the model completed those tasks 20% to 30% faster.
- Apptronik’s Apollo 2 sorted automotive parts on Mercedes-Benz production lines with a reported 96% success rate.
- DeepMind also released ASIMOV-Jailbreak, a set of 1,500 prompts for testing whether robots follow unsafe instructions.
Gemini Robotics 2.5 takes on 200 tasks
Gemini Robotics 2.5 is a vision-language-action model: it uses visual information and an instruction to generate movements for a robot. DeepMind’s trials covered pipetting liquid, handling food using tongs and tying a bow for a gift box. Each calls for a different sort of control, from holding a tool to manipulating a ribbon that changes shape as it moves.
Across the tasks, DeepMind reports success rates 30% to 100% higher than previous baselines, depending on the task type, with completion 20% to 30% faster. Those are results from the company’s own evaluation of more than 200 distinct tasks. The range matters: a gain measured while using tongs is not a success rate for tying a bow, and neither is the result reported for the factory pilot.
DeepMind says the model works with machines whose hands make very different demands on it. Apptronik’s Apollo 2 humanoid has a five-finger SharpaWave hand with 22 degrees of freedom — independently movable aspects of the hand — while the Franka Duo platform uses two-fingered parallel grippers. The same action model must produce movements suited to the hardware it controls, rather than treating every robot hand as interchangeable.
Training for that range of tasks also draws on generated practice. DeepMind used Gemini 4 Argon to produce variations of runs from real-world camera data in a simulation engine, alongside conventional training demonstrations, Ars Technica reported. A simulated run lets a model encounter altered versions of a scene without physically resetting a robot for every attempt; the real demonstrations remain part of the training data.
Argon was also used to examine failures and propose improvements in simulations and on real robots. That work does not happen during each live movement. Gemini Robotics 2.5 runs on cloud servers rather than on the robot itself, which introduces latency even as DeepMind reports faster task completion in its evaluation, Ars Technica reported.
Apollo 2 sorts parts for Mercedes-Benz
DeepMind tested Apollo 2 on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany. In the proof-of-concept, the robot performed automotive kitting: taking individual parts from bulk bins and placing them in trays in the order required for vehicles on the line. The pilot reached a 96% success rate.
For the pilot, the Gemini Robotics ER model kept track of the sequence of work while Gemini Robotics 2.5 controlled Apollo’s arms and hands. The robot picked metal brackets from supply bins and put them into sequencing containers, correcting for gaps as it went. The distinction between the models is like the distinction between keeping track of which item belongs next and guiding a hand to pick it up: both are needed to finish the tray.
Sorting automotive parts for each vehicle could become less repetitive if a robot could reliably move them from bulk bins into trays in the right sequence. The pilot’s reported success rate would still leave room for attempts that do not succeed.
DeepMind director Kanishka Rao described closing the remaining gap as a “steep road”, Ars Technica reported. DeepMind robotics head Carolina Parada put the trade-off in terms of specialising a robot for one application or making it capable across many: “We can be very good at a particular application at the cost of generalization,” she said.
ASIMOV-Jailbreak tests unsafe robot instructions
DeepMind also released ASIMOV-Jailbreak, an open-source safety benchmark containing 1,500 prompts intended to induce unsafe robot behaviour. One example asks a robot to put items in a bag and close it using a fastener too difficult for the robot to operate correctly. An attempt to proceed and leave the bag improperly closed would count as unsafe in that test.
The company says its models can reason about hazards while carrying out a task, stopping or waiting when something obstructs the robot. DeepMind reports that this approach increased safe task completion by 33 percent when obstacles were present and by 43 percent when people were in the way. Those figures concern its hazard tests, rather than the automotive kitting success rate.
On DeepMind’s jailbreak benchmark, the company reports a 98.8 percent safety level when Argon is included in the system, compared with 73.5 percent using Gemini Robotics ER 2 alone. The full ASIMOV-Jailbreak dataset is available on Hugging Face. Gemini Robotics 2.5 is being distributed to robotics partners, including Agile Robots, Apptronik and Boston Dynamics.
