Google DeepMind unveils Gemini Robotics 2.5, hitting 96% success sorting parts at Mercedes-Benz plants

Google DeepMind has released Gemini Robotics 2.5, a new AI model for robots that it says can learn hundreds of everyday and industrial skills, from folding laundry and serving pastries to sorting car parts on a factory floor.

According to a DeepMind blog post on Nov 10, the vision-language-action (VLA) model was tested on more than 200 tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid. Compared with earlier baselines, it achieved 30% to 100% higher success rates depending on the task and finished tasks 20% to 30% faster. It is rolling out to DeepMind's robotics partners first.

DeepMind also shared results from a real factory pilot. Apptronik's Apollo 2 humanoid, running the new model, picked metal brackets from bulk bins and packed them in build order on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, with a 96% success rate. "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," said Apptronik CEO Jeff Cardenas.

On safety, DeepMind said the model completed tasks safely 33% more often around obstacles and 43% more often near people. It also released ASIMOV-Jailbreak, an open-source benchmark of prompts that disguise unsafe requests to robots; Gemini 4 Argon scored 99.5% jailbreak resistance, up from 82% for the previous baseline. DeepMind's announcement on X drew about 1,600 likes in just over an hour.

Gemini 4 Argon helped build the model by acting as a simulation engine that turns a single real camera scene into thousands of varied practice runs. DeepMind gave robots, including humanoids, whole-body control with Gemini Robotics 2 in July.