Google DeepMind announces Gemini Robotics 2.5 and a kitting pilot with Apptronik and Mercedes-Benz
DeepMind says its new vision-language-action model raises success rates by 30% to 100% over earlier baselines, and reports a 96% success rate for Apollo 2 humanoids kitting parts in two car plants.
Google DeepMind announced Gemini Robotics 2.5, its latest vision-language-action (VLA) model, on 10 November 2026. DeepMind says it evaluated the model on more than 200 tasks, including serving food with tongs, tying a bow and pipetting. According to DeepMind, success rates are 30% to 100% higher than previous baselines depending on the task type, and tasks finish 20% to 30% faster. The post does not name the baselines or give per-task numbers.
The model runs on several robot types. The post names Apptronik’s Apollo 2 humanoid, fitted with the SharpaWave hand (five fingers, 22 DoF), and a Franka Duo setup with two-finger parallel grippers. It does not give Apollo 2’s payload or reach.
The industrial test was a proof of concept in two Mercedes-Benz plants, at Kecskemét in Hungary and Sindelfingen in Germany, on live lines. Apollo 2 robots did kitting: taking metal brackets from bulk bins and placing them in sequencing trays in the order each car needs. An embodied reasoning model tracked the workflow while the 2.5 action model drove the arms and hands. DeepMind reports a 96% success rate and calls the robots autonomous in this task.
On safety, DeepMind says inference-time guidance raised safe task completion by 33% around obstacles and 43% around people in its evaluations. It is also releasing ASIMOV-Jailbreak, an open benchmark of prompts that disguise unsafe requests.
Why it matters
Bin-to-tray kitting is a real line-side job with a clear pass or fail, which makes it a better test than a lab demo. A 96% success rate still leaves 4% of attempts failing, and the post does not say what counted as an attempt. Ask how failures are detected and recovered before a cell like this goes near takt time. All numbers here are DeepMind’s own, from a launch post. The post does not say whether or when developers outside its partners can use the model.