Google DeepMind releases Gemini Robotics 2.5, reports 96% success in Mercedes-Benz factory pilot
Google DeepMind has introduced Gemini Robotics 2.5, a vision-language-action (VLA) model that turns camera images and spoken or written instructions into robot movements. The company says a humanoid robot running it reached a 96 percent success rate in a pilot on active Mercedes-Benz production lines.
DeepMind says it evaluated the model on more than 200 tasks, among them serving food with tongs and pipetting liquids. Compared with the company's own earlier baselines, it reports success rates that are 30 to 100 percent higher, depending on the kind of task, and completion times that are 20 to 30 percent shorter. Ars Technica reported that Apptronik's Apollo 2 humanoid with the new model can also tie a bow, a difficult job for robots because ribbon is deformable.
Pilot on Mercedes-Benz production lines
DeepMind worked with Apptronik and Mercedes-Benz on the test, which ran in Kecskemét, Hungary, and Sindelfingen, Germany, according to The Decoder. Apollo 2 handled kitting, picking metal components out of loose bins and arranging them in trays following the sequence each vehicle on the line needs. One model, Gemini Robotics ER, supervised the workflow while Gemini Robotics 2.5 drove the hands and arms.
Apptronik CEO Jeff Cardenas said the models got markedly better after a few weeks of post-training and carried over to Mercedes production with little loss in performance, The Decoder reported. Ars Technica noted that 96 percent is probably still short of what a manufacturer needs, and quoted DeepMind director Kanishka Rao as saying that closing the remaining gap will be a "steep road."
DeepMind Robotics head Carolina Parada described the trade-off in building general-purpose robots.
“We can be very good at a particular application at the cost of generalization. Or we can just push on the generalization axis and maybe not get to 99.9 in every application.”
Simulation for training data
Robot training data is scarcer than the text used for chatbots. DeepMind used its Gemini 4 Argon language model as a simulation engine, according to both reports. Argon can generate thousands of simulated runs from a real camera scene, creating a digital twin, and the company combined this with real-world demonstrations to help the robot adapt to new scenes and objects.
Ars Technica reported that the 2.5 model still runs on cloud-based servers, so gains in accuracy and speed come with latency.
Safety benchmark
DeepMind says the model uses real-time reasoning to avoid hazards, stopping or waiting when something blocks its path. In the company's evaluations, safe task completion rose 33 percent near obstacles and 43 percent near people.
DeepMind also released ASIMOV-Jailbreak on Hugging Face, an open-source benchmark of prompts that disguise unsafe requests to test whether robot models can be tricked into dangerous actions. The two outlets differ on its size and results: Ars Technica cites 1,500 prompts and a 98.8 percent safety level with Argon, against 73.5 percent for Gemini Robotics ER 2. The Decoder cites hundreds of prompts and 99.5 percent versus 82 percent.
Gemini Robotics 2.5 is not generally available. It is being rolled out to robotics partners; Ars Technica lists Agile Robots, Apptronik and Boston Dynamics, while The Decoder names Agile, Apptronik and Mercedes-Benz.
