--- title: Google DeepMind releases Gemini Robotics 2.5 and Gemini... description: Google DeepMind released Gemini Robotics 2.5, its most capable robotics model so far, alongside a separate physical-reasoning model called Gemini 4 Argon.... canonical: https://daily.dev/posts/google-deepmind-releases-gemini-robotics-2-5-and-gemini-4-argon-for-robot-control-ltagefdzj twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Google DeepMind releases Gemini Robotics 2.5 and Gemini 4 Argon for robot control | daily.dev og:description: Google DeepMind released Gemini Robotics 2.5, its most capable robotics model so far, alongside a separate physical-reasoning model called Gemini 4 Argon.... og:url: https://daily.dev/posts/google-deepmind-releases-gemini-robotics-2-5-and-gemini-4-argon-for-robot-control-ltagefdzj og:image: https://api.daily.dev/og/posts/lTagEFdzJ.png og:image:alt: Google DeepMind releases Gemini Robotics 2.5 and Gemini 4 Argon for robot control og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Google DeepMind releases Gemini Robotics 2.5 and Gemini 4 Argon for robot control **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Google DeepMind released Gemini Robotics 2.5, its most capable robotics model so far, alongside a separate physical-reasoning model called Gemini 4 Argon. Gemini Robotics 2.5 reports success rates 30 to 100% higher than prior baselines and runs 20 to 30% faster. In a proof of concept with Apptronik and Mercedes-Benz, an Apollo 2 humanoid picked and organized car parts with 96% success. Argon directed a Boston Dynamics Spot robot and a simple arm, scored 87% on the ERQA benchmark, and helped raise a policy's real-world success rate from 56% to 75% after training on simulated episodes generated from one photo. DeepMind also open-sourced a safety benchmark called ASIMOV-Jailbreak. Argon is available to Google AI Ultra subscribers and via the paid Gemini API as gemini-4-argon. ## Content Google has put out two robotics models: Gemini Robotics 2.5 from DeepMind, a vision-language-action model for controlling robots, and Gemini 4 Argon, a general-purpose physical-reasoning model. Robotics 2.5 was built with support from Argon. ## Gemini Robotics 2.5 DeepMind calls it its most capable robotics model yet. It was evaluated on more than 200 distinct tasks, including serving food with tongs and pipetting liquid. Against earlier baselines it posts 30% to 100% higher success rates and finishes tasks 20% to 30% faster. The headline demo is an industrial one. In a proof of concept with Apptronik and Mercedes-Benz, the Apollo 2 humanoid picked and packed car parts for assembly, a kitting task, with a 96% success rate. The pilot ran at Mercedes-Benz plants in Hungary and Germany. Safety numbers are part of the release. Safety guidance improved safe task completion by 33% around obstacles and 43% around people. DeepMind also open-sourced ASIMOV-Jailbreak, a benchmark for robotics-specific jailbreak resistance. Gemini 4 Argon scored 99.5% on it, against 82% for the prior baseline. Partners get Robotics 2.5 first. ## Gemini 4 Argon Argon covers robot control, simulation, and policy improvement. Google shows it directing a Boston Dynamics Spot through a foot-pedal task after the robot practiced in a MuJoCo digital twin. It also steers a simpler arm using visual coordinates and HTTP motion primitives. The policy-improvement result is the most concrete. Google fine-tuned a vision-language-action policy on synthesized, domain-randomized MuJoCo episodes generated from a single RGB photo. Success on a real setup rose from 56% to 75%. On the ERQA benchmark, Argon scores 87% using pure reasoning, which Google says leads other frontier models on that test. Argon is available to Google AI Ultra subscribers and through the paid Gemini API as gemini-4-argon. ## Worth keeping in mind The 96% figure comes from a single pilot on one task type, reported by Google and its partners. Still, going from lab demos like tongs and pipettes to a car-factory line is the interesting part, and the safety benchmark being open means others can check the claims. ## Questions this post answers ### What is Gemini Robotics 2.5 and how does it perform compared to previous models? Gemini Robotics 2.5 is Google DeepMind's physical AI model for robot control, described as its most capable robotics model so far. It reports success rates 30 to 100% higher than prior baselines and runs 20 to 30% faster. In a proof of concept, an Apptronik Apollo 2 humanoid used it to pick and organize car parts for assembly with 96% success at Mercedes-Benz plants. _Teams evaluating physical AI models for robot control can track releases like this on daily.dev._ ### What is Gemini 4 Argon and how is it accessed? Gemini 4 Argon is a general-purpose physical-reasoning model from Google used for robot control, simulation, and policy improvement, built to support Gemini Robotics 2.5. It scores 87% on the ERQA benchmark using pure reasoning. It is available to Google AI Ultra subscribers and through the paid Gemini API under the name gemini-4-argon. _Developers weighing access and pricing for new physical-reasoning APIs can follow updates on daily.dev._ ### Can a robot policy be improved using just a single photo instead of real-world trials? Yes. Starting from a single RGB photo, Gemini 4 Argon synthesized domain-randomized MuJoCo simulation episodes and used them to fine-tune a vision-language-action policy. Success on a real robotic setup rose from 56% to 75% after this simulation-based fine-tuning, without additional real-world data collection. _Anyone exploring sim-to-real training shortcuts for robot policies can follow this approach on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 14 comments across x (as of 2026-11-10). **TL;DR:** Reactions focus mostly on the 96% car-parts sorting demo, with curiosity about what the remaining failures look like and questions about how the benchmark numbers were actually measured. **Sentiment:** 30% positive · 45% mixed · 25% skeptical **The case for** - Several find the car-parts pilot exciting as a concrete, useful real-world robotics task. - Some note it's reassuring that the tech is tested on real factory parts rather than a lab demo. **The pushback** - Multiple commenters question what's behind the 4% failure rate and whether it includes recovery after a miss. - One points out the model still runs on cloud servers rather than on the robot itself, and access is limited to select partners. - Another argues the real test is generalization to new parts and bins, which is where such pilots usually break. **By community** - x (mixed): Interest in the car-parts sorting demo is tempered by pointed questions about failure rates, measurement methodology, and generalization beyond the pilot. **Open questions** - What exactly counts in the 96% success rate denominator, and does it include recovery after misses? - How well does the system generalize to new part shapes and bins outside the pilot setup? - When does this move from a proof-of-concept pilot to actual factory floor deployment? **Highlights** > @testingcatalog What’s the denominator behind 96%—how many picks, and did that include recovery after a miss? Those details make the pilot much easier to compare. > — [pwndstar on x](https://x.com/pwndstar/status/2120178469799960824) > @testingcatalog Robots doing the boring parts > — [xabzxbt on x](https://x.com/xabzxbt/status/2120180129934483529) > @testingcatalog The last 4% is where a bolt ends up in the cupholder. Curious how it fails on a real line vs. a POC. > — [cunlegetu on x](https://x.com/cunlegetu/status/2120176166669840436) > @testingcatalog 96% on sorting car parts sounds great until you ask who catches the other 4% > — [lozoboss on x](https://x.com/lozoboss/status/2120177169452470407) **Source threads** - [x](https://x.com/cortexwire/status/2120179109678457234) · 0 points · 0 comments - [x](https://x.com/window98se/status/2120180763454996822) · 0 points · 0 comments - [x](https://x.com/testingcatalog/status/2120173652641476658) · 0 points · 14 comments ## Similar posts on daily.dev - [Gemini 4 Argon: Google’s New AI Is Built to Think and Work for Much Longer](https://daily.dev/posts/gemini-4-argon-google-s-new-ai-is-built-to-think-and-work-for-much-longer-ibbgfhpdj) · Medium · 0 upvotes · 0 comments --- Tags: [#robotics](https://daily.dev/tags/robotics), [#google-deepmind](https://daily.dev/tags/google-deepmind) [View this post on daily.dev](https://daily.dev/posts/google-deepmind-releases-gemini-robotics-2-5-and-gemini-4-argon-for-robot-control-ltagefdzj) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Google DeepMind releases Gemini Robotics 2.5 and Gemini 4 Argon for robot control","url":"https://daily.dev/posts/google-deepmind-releases-gemini-robotics-2-5-and-gemini-4-argon-for-robot-control-ltagefdzj","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/google-deepmind-releases-gemini-robotics-2-5-and-gemini-4-argon-for-robot-control-ltagefdzj"},"datePublished":"2026-11-10T15:49:16.956Z","dateModified":"2026-11-10T16:09:18.808Z","description":"Google DeepMind released Gemini Robotics 2.5, its most capable robotics model so far, alongside a separate physical-reasoning model called Gemini 4 Argon....","image":"https://pbs.twimg.com/media/HWxlJmua0AA4Kre.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HWxlJmua0AA4Kre.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/google-deepmind-releases-gemini-robotics-2-5-and-gemini-4-argon-for-robot-control-ltagefdzj","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"robotics,google-deepmind","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Google DeepMind releases Gemini Robotics 2.5 and Gemini 4 Argon for robot control"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/google-deepmind-releases-gemini-robotics-2-5-and-gemini-4-argon-for-robot-control-ltagefdzj#faq","mainEntity":[{"@type":"Question","name":"What is Gemini Robotics 2.5 and how does it perform compared to previous models?","acceptedAnswer":{"@type":"Answer","text":"Gemini Robotics 2.5 is Google DeepMind's physical AI model for robot control, described as its most capable robotics model so far. It reports success rates 30 to 100% higher than prior baselines and runs 20 to 30% faster. In a proof of concept, an Apptronik Apollo 2 humanoid used it to pick and organize car parts for assembly with 96% success at Mercedes-Benz plants. Teams evaluating physical AI models for robot control can track releases like this on daily.dev."}},{"@type":"Question","name":"What is Gemini 4 Argon and how is it accessed?","acceptedAnswer":{"@type":"Answer","text":"Gemini 4 Argon is a general-purpose physical-reasoning model from Google used for robot control, simulation, and policy improvement, built to support Gemini Robotics 2.5. It scores 87% on the ERQA benchmark using pure reasoning. It is available to Google AI Ultra subscribers and through the paid Gemini API under the name gemini-4-argon. Developers weighing access and pricing for new physical-reasoning APIs can follow updates on daily.dev."}},{"@type":"Question","name":"Can a robot policy be improved using just a single photo instead of real-world trials?","acceptedAnswer":{"@type":"Answer","text":"Yes. Starting from a single RGB photo, Gemini 4 Argon synthesized domain-randomized MuJoCo simulation episodes and used them to fine-tune a vision-language-action policy. Success on a real robotic setup rose from 56% to 75% after this simulation-based fine-tuning, without additional real-world data collection. Anyone exploring sim-to-real training shortcuts for robot policies can follow this approach on daily.dev."}}]} ```