New version of robotics model improves dexterity, reliability and safety

Google DeepMind, Google's research division, on Nov. 10 (Nov. 11 Japan time) announced "Gemini Robotics 2.5," an AI model that controls robots. It is a "VLA (vision-language-action) model" that takes camera images and instructions in words and turns them into robot actions, and the company positions it as its most capable model of this kind to date. The company said it improved 3 areas, dexterity, reliability in real-world settings and safety, and that it tested the model on more than 200 kinds of tasks.

Gemini Robotics 2.5 is here. πŸ€– Our latest physical AI is able to learn hundreds of everyday skills - from folding laundry and serving pastries to autonomously sorting car parts on the factory floor. Here’s a snapshot of its capabilities 🧡

@GoogleDeepMind Β· November 11, 2026

The previous version, "Gemini Robotics 2," was announced in July and enabled whole-body control of humanoid robots. At that time, "Gemini Robotics ER 2," which plans tasks, was released in the development environment "Google AI Studio," while the model that generates actions was limited to early-access partners. The new model was developed using "Gemini 4 Argon," released in October. The company said it generated thousands of virtual practice runs from real camera footage and trained the model on them combined with demonstrations of real tasks.

Success rates up from previous version, but folding pants stays at 40%

According to Google DeepMind, the model can now handle fine tasks such as serving food with tongs, tying a ribbon on a gift box and measuring liquid with a pipette. Compared with previous baselines, the company said success rates were 30% to 100% higher depending on the task, and the time needed to complete tasks was 20% to 30% shorter.

In a chart of fine manual tasks on the official blog, the success rate for packing pastries with tongs rose from 5% with the previous version to 100%. Untying a ribbon went from 0% to 70%, and hanging a cardigan from 0% to 62%. Folding pants rose from 0% to 40%, but more than half of the attempts still failed. All are the company's own measurements, and conditions such as the number of trials have not been disclosed.

Chart of success rates for fine manual tasks published by Google DeepMind on its official blog, comparing Gemini Robotics 2 (hatched) and Gemini Robotics 2.5 (blue) for each task
Chart of success rates for fine manual tasks published by Google DeepMind on its official blog, comparing Gemini Robotics 2 (hatched) and Gemini Robotics 2.5 (blue) for each task (Source: Google DeepMind: Gemini Robotics 2.5: Accelerating Physical Intelligence (Nov. 10))

The supported robots named were Apptronik's humanoid robot "Apollo 2" and the two-armed robot "Franka Duo." Apollo 2 is fitted with five-fingered hands with 22 degrees of freedom, while Franka Duo uses two-fingered grippers.

Success rates for fine manual tasks (%). Google DeepMind's measurements. Values from the descriptive text (alt text) accompanying the chart on the official blog
  • Gemini Robotics 2
  • Gemini Robotics 2.5
  • Packing pastries with tongsGoogle5 β†’ 100
  • Untying a ribbonGoogle0 β†’ 70
  • Hanging a cardiganGoogle0 β†’ 62
  • Folding pantsGoogle0 β†’ 40

96% success rate in Mercedes production line trial

Google DeepMind worked with Apptronik and Mercedes-Benz on trials on operating production lines in Kecskemet, Hungary, and Sindelfingen, Germany. Apollo 2 took car parts packed together in boxes and loaded them into trays in the order needed for each vehicle. "Gemini Robotics ER" handled understanding the workflow, while Gemini Robotics 2.5 controlled the hand and arm movements.

The success rate was 96%. The company described the trial as a proof of concept and "promising progress" toward broad use in factories and other settings. Apptronik CEO Jeff Cardenas said the model improved with a few weeks of additional training and that its performance dropped only slightly when moved to Mercedes-Benz's production floor.

The duration of the trial and the number of tasks performed have not been disclosed. No third-party evaluation could be found as of Nov. 12.

Evaluation released for avoiding hazards and resisting misuse of instructions

On safety, the company added a mechanism that guides movements with safety-related reasoning at the point actions are decided. According to Google DeepMind, the rate of safely completing tasks near obstacles improved by 33%, and by 43% in situations with people nearby.

The company also released "ASIMOV-Jailbreak," an evaluation that measures resistance to instructions trying to make robots act dangerously, as open source on the AI sharing site Hugging Face. It tests hundreds of instructions that disguise dangerous requests through rewording, across dozens of scenarios. On this evaluation, Gemini 4 Argon rejected 99.5% of the instructions. "Gemini Robotics ER 2," used as the previous baseline, rejected 82%.

Offered first to robotics partners, no API or pricing disclosed

Gemini Robotics 2.5 will be offered to partners in the robotics field first. The blog named Agile, Apptronik and Mercedes-Benz as partners. The model page has a sign-up for an early-access waitlist.

The official blog and model page do not mention API availability, pricing, regions or timing for general availability. No model card or technical report for Gemini Robotics 2.5 itself could be found as of Nov. 12.