--- title: "Google DeepMind releases Gemini Robotics 2.5 to robotics partners with new safety benchmark" url: https://llmreport.gr/news/google-deepmind-releases-gemini-robotics-2-5-to-robotics-partners-with-new/ language: en published: 2026-11-10T23:12:54.008Z updated: 2026-11-10T23:12:54.013Z site: "LLM Report" section: "Models" --- # Google DeepMind releases Gemini Robotics 2.5 to robotics partners with new safety benchmark Google DeepMind says its Gemini Robotics 2.5 model lifts task success rates by 30% to 100% and reached 96% in a Mercedes-Benz factory pilot. It is limited to robotics partners for now. > Written with AI from the sources listed below. Editorial responsibility: Konstantinos Perros. How we write: https://llmreport.gr/how-we-write/ ## Key facts - Google DeepMind introduced Gemini Robotics 2.5, which it calls its most capable vision-language-action model, on 10 November 2026; it is rolling out to robotics partners rather than being generally available. [1] [2] - Google DeepMind reports a 30% to 100% higher success rate than previous baselines, depending on task type, and tasks completed 20% to 30% faster. [1] [2] - In a pilot on Mercedes-Benz production lines, Apptronik's Apollo 2 humanoid picked and packed automotive parts with a 96% success rate. [1] [2] - DeepMind released ASIMOV-Jailbreak, an open-source benchmark of adversarial prompts for robots; the two sources give different prompt counts and scores for it. [1] [2] ## What DeepMind released Google DeepMind has introduced Gemini Robotics 2.5, which it describes as its most capable vision-language-action (VLA) model so far. A VLA model turns instructions and visual input into physical robot actions. According to Ars Technica, the platform has two main layers: Gemini Robotics ER, an embodied reasoning model that uses live video to plan how a task should be done, and the Gemini Robotics VLA, which generates the movements and is now at version 2.5. DeepMind says the model widens the range of skills a robot can learn and shortens the time it takes to learn them. It was evaluated on more than 200 tasks, including serving food with tongs, tying a bow on a gift box and pipetting a liquid. Compared with earlier baselines, DeepMind reports a 30% to 100% higher success rate depending on the task, and completion 20% to 30% faster. Ars Technica notes that the bow-tying result is significant because deformable objects such as ribbon are hard for robots to handle. The model works across several robot types. DeepMind names Apptronik's Apollo 2 humanoid, whose five-fingered SharpaWave hand has 22 degrees of freedom, and the Franka Duo platform, which uses a pair of parallel grippers with two fingers each. ## Factory pilot with Apptronik and Mercedes-Benz To test the model outside the lab, DeepMind worked with Apptronik and Mercedes-Benz on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 robots carried out automotive kitting: taking parts from bulk bins and placing them in trays in the order each vehicle needs. In DeepMind's description, the embodied reasoning model tracked the workflow while the 2.5 actions model controlled the hands and arms, and the robot handled metal brackets with a 96% success rate. Ars Technica puts the result in perspective. It says 96% is strong compared with earlier robotic AI systems but probably short of what a manufacturer needs, since a person doing repetitive work would get it right almost every time. DeepMind director Kanishka Rao told the outlet that closing the remaining 4% would be a steep road. DeepMind Robotics head Carolina Parada described a trade-off: a system can be tuned for one application at the expense of generalisation, or it can push for generality and may not reach near-perfect reliability in every setting. ## Safety features and the ASIMOV-Jailbreak benchmark DeepMind says Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards. In its evaluations, safe task completion improved by 33% when obstacles were present and by 43% when people were nearby. Ars Technica adds that the robot stops or waits until the way is clear. DeepMind is also releasing ASIMOV-Jailbreak, an open-source benchmark of prompts that disguise unsafe requests, aimed at testing robots against jailbreaks. The sources differ on the details. DeepMind says the set holds hundreds of prompts across dozens of scenes and reports that Gemini 4 Argon reaches 99.5% jailbreak resistance, against 82% for Gemini Robotics ER 2. Ars Technica reports 1,500 prompts, with Argon in the loop at a 98.8% safety level and Gemini Robotics ER 2 alone at 73.5%. Ars says the dataset is on Hugging Face. Ars Technica also reports that no agreed method exists for testing physical AI for safety, and that DeepMind hopes this benchmark moves the industry toward one. ## Role of Gemini 4 Argon, limits and availability DeepMind credits Gemini 4 Argon with speeding up development by acting as a simulation engine. Argon can generate thousands of varied practice runs from one real camera scene, building a digital twin, and when combined with real training demonstrations this helps the robot generalise to new scenes and objects. It can also suggest why failures occurred from video and help test fixes. Ars Technica reports that this analysis does not run in real time and that the 2.5 release still runs on cloud servers, so latency remains despite the gains in accuracy and speed. Gemini Robotics 2.5 is not generally available. DeepMind says it is rolling out to robotics partners, and Ars lists Agile Robots, Apptronik and Boston Dynamics among those with access. ## Questions and answers **Can developers use Gemini Robotics 2.5 now?** Not generally. DeepMind says it is rolling out to its robotics partners, and Ars Technica names Agile Robots, Apptronik and Boston Dynamics among those with access. **Where was the 96% success rate measured?** On active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, where Apptronik's Apollo 2 humanoid sorted automotive parts into trays. **Where can the ASIMOV-Jailbreak benchmark be found?** It is open source. Ars Technica says the dataset is available on Hugging Face. ## Sources 1. Google DeepMind: [Gemini Robotics 2.5 accelerates real-world general dexterity](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence) 2. Ars Technica: [DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak"](https://arstechnica.com/google/2026/11/deepmind-reveals-gemini-robotics-2-5-and-open-safety-benchmark-asimov-jailbreak)