Google DeepMind’s new robot brain has humanoids packing car parts at Mercedes-Benz

Google DeepMind launched Gemini Robotics 2.5 on Tuesday, calling it its "most capable vision-language-action (VLA) model to date," and reported that Apptronik Apollo 2 humanoids running it picked and packed parts on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, with a 96% success rate. DeepMind said Gemini Robotics 2.5 succeeds 30% to 100% more often and finishes tasks 20% to 30% faster than its previous models across more than 200 tested tasks, though the lab noted there is no independent evaluation yet. DeepMind also released ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts, on which Gemini 4 Argon resisted 99.5% of jailbreak attempts versus 82% for Gemini Robotics ER 2.

Apptronik’s Apollo 2 humanoids sorting parts on a Mercedes-Benz production line in Google DeepMind’s launch images for Gemini Robotics 2.5. Image: Google DeepMind Google DeepMind’s newest robot brain has been sorting car parts on Mercedes-Benz production lines. The lab launched Gemini Robotics 2.5 https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/ on Tuesday, calling it its “most capable vision-language-action VLA model to date”, and said Apptronik’s Apollo 2 humanoids running it picked and packed parts at two Mercedes factories with a 96% success rate. What Gemini Robotics 2.5 can do A VLA model takes what a robot sees and what it is told, and turns that into motor commands for its arms and hands. DeepMind says the new version widens the range of skills a robot can learn and speeds up how fast it picks them up. It follows Gemini Robotics 2, which brought what the lab calls “whole-body intelligence” to humanoids in July. The lab tested it on more than 200 distinct tasks, such as serving food with tongs, tying a bow on a gift box and pipetting a liquid. Compared with its previous models, DeepMind says Gemini Robotics 2.5 succeeds 30% to 100% more often , depending on the type of task, and finishes tasks 20% to 30% faster. All of these are DeepMind’s own results; there is no independent evaluation yet. Like earlier versions, it works across different robot bodies. DeepMind names Apptronik’s Apollo 2 humanoid, fitted with the five-finger SharpaWave hand 22 degrees of freedom , and the Franka Duo, a two-armed platform with simple two-fingered grippers. Humanoids on the Mercedes-Benz line With Apptronik and Mercedes-Benz, DeepMind put Apollo 2 robots on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany, to do “kitting”: taking individual parts from bulk bins and laying them out in trays in the exact order each car on the line needs them. Two models split the work. DeepMind’s embodied reasoning model, Gemini Robotics ER https://deepmind.google/models/gemini-robotics/embodied-reasoning/ , keeps track of the job, while Gemini Robotics 2.5 controls the hands and arms as the robot reaches into bins, picks up metal brackets and places them in sequencing containers, “all while correcting for any gaps”. The pilot hit 96%. DeepMind calls it a proof of concept, and neither company has said whether humanoids will stay on the lines. Apptronik’s chief executive said in the post: Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off. Jeff Cardenas, co-founder and CEO, Apptronik DeepMind pitches kitting as “monotonous” work that frees technicians for complex build tasks. Not every factory worker sees robots that way: some at Tesla have balked at training the machines https://madrobot.blog/2026/09/26/tesla-optimus-workers-balk-training-robot-replacements/ built to replace them. A jailbreak test built for robots A chatbot tricked into saying something harmful is one problem; a robot tricked into doing something harmful is another. DeepMind is releasing ASIMOV-Jailbreak https://huggingface.co/datasets/google/asimov agentic , an open-source benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes. On it, Gemini 4 Argon resisted 99.5% of jailbreak attempts, against 82% for the previous baseline, Gemini Robotics ER 2. DeepMind says Argon also leads its agentic robotics safety evaluations https://storage.googleapis.com/deepmind-media/gemini/gemini 4 safety report.pdf page=200 on refusing unsafe tasks, resolving uncertainty and working close to people. Gemini Robotics 2.5 itself gets what DeepMind calls inference-time reasoning guidance, which steers the robot around physical hazards as it moves. In the lab’s tests, that raised safe task completion by 33% around obstacles and by 43% around nearby people. Built with help from Gemini 4 Argon DeepMind credits Gemini 4 Argon https://madrobot.blog/2026/10/13/google-gemini-4-argon-available-ai-pro-ultra-api-robotics/ , its most capable model, with much of the speed-up. Argon acted as a simulation engine: from a single real camera scene it can generate thousands of varied practice runs and build a realistic digital twin, which, mixed with real demonstrations, helps the robot cope with new scenes and objects. It can also work out from video why attempts failed and test fixes. It is the same push behind Genie 4 https://madrobot.blog/2026/10/27/google-deepmind-genie-4-world-model-sound-dialogue/ , the lab’s world model, which DeepMind also pitches for building training worlds for robots. Who can use it Gemini Robotics 2.5 is “rolling out to our robotics partners”. DeepMind’s Gemini Robotics page https://deepmind.google/models/gemini-robotics/ lists Agile Robots, Apptronik and Boston Dynamics https://madrobot.blog/2026/10/06/boston-dynamics-rohit-prasad-ceo-amazon-alexa-agi-physical-ai/ as partners and says more than 100 trusted testers, from industrial automation firms to robotics start-ups, are working with its models. Others can join a waitlist for early access. DeepMind hasn’t announced a public API, a price or a rollout by country, so outside that circle it isn’t available anywhere yet. Why it matters Humanoid robots have spent years in demo videos; this is one of the clearest claims yet of one doing useful, repetitive work on a real car line at a high success rate. A pilot is not a deployment, but a robot-specific jailbreak test released in the open gives others a way to check how safe these machines are before they work alongside people. Sources: Google DeepMind: Gemini Robotics 2.5 accelerates real-world general dexterity https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/ ; Google DeepMind: Gemini Robotics https://deepmind.google/models/gemini-robotics/ ; ASIMOV-Jailbreak benchmark Hugging Face https://huggingface.co/datasets/google/asimov agentic ; Gemini 4 safety report https://storage.googleapis.com/deepmind-media/gemini/gemini 4 safety report.pdf page=200 .