Google DeepMind introduces Gemini Robotics 2.5, its robot-control model, and says Apptronik's Apollo 2 humanoid reached 96% success sorting parts in a Mercedes-Benz pilot

Google DeepMind introduced Gemini Robotics 2.5 on 10 November, calling it “our most capable vision-language-action (VLA) model to date.” A VLA model turns what a robot sees and is told into movements. DeepMind says it tested the model on more than 200 distinct tasks, and claims success rates 30% to 100% higher than its earlier models, depending on the kind of task, with work done 20% to 30% faster.

The main example is a pilot with Apptronik's Apollo 2 humanoid doing automotive kitting, sorting parts into trays, at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany. “Achieving a 96% success rate, this pilot shows promising progress toward widespread real-world industrial applications,” the post says.

On safety, DeepMind says reasoning while the robot works raised safe task completion by 33% around obstacles and by 43% around nearby people. It also released ASIMOV-Jailbreak, an open benchmark of disguised unsafe requests, and says Gemini 4 Argon, the Google model behind the work, reaches 99.5% jailbreak resistance against 82% for its previous baseline, Gemini Robotics ER 2. These are DeepMind's own results.

“Gemini Robotics 2.5 is rolling out to our robotics partners,” the post says. It gives no general release date, API or price.

• Claimed Tested on more than 200 tasks, with 30% to 100% higher success and 20% to 30% faster work than DeepMind's earlier models. Google DeepMind
• Claimed Apptronik's Apollo 2 reached a 96% success rate on parts kitting at Mercedes-Benz plants in Kecskemét and Sindelfingen. Google DeepMind
• Confirmed The model is rolling out to robotics partners; the page names no public release, API or price. Google DeepMind