Gemini Robotics 2.5 hits 96% success on parts handling in a Mercedes-Benz factory test

Robotics foundation models are moving from demo videos to factory numbers. DeepMind presented Gemini Robotics 2.5 at the Conference on Robot Learning in Austin. Its headline result is 96% success on automotive parts handling in a Mercedes-Benz plant using Apptronik's Apollo 2 humanoid. Across an evaluation of more than 200 tasks, DeepMind reports gains of 30% to 100% over earlier baselines.

Gemini Robotics is a vision-language-action (VLA) model. It takes camera input and language instructions and outputs motor actions directly, inheriting world knowledge from the Gemini backbone. That lets one model generalize across tasks instead of requiring a hand-engineered pipeline per part or station. Apptronik supplies the hardware, and Mercedes-Benz provides a real production environment rather than a lab mock-up.

The field is crowded. Tesla's Optimus, Figure, Physical Intelligence and NVIDIA's GR00T all chase the same promise, and NVIDIA and others publish open robotics and tactile models on Hugging Face, such as LingBot-Touch this week. DeepMind's advantage is the Gemini backbone and a clear industrial partner. Coverage from SiliconANGLE notes DeepMind released a next-generation gene prediction model at the same time, which reflects Google's broad push across applied AI.

Skeptically, 96% on a specific parts-handling task is impressive, but factories typically need much higher reliability before removing human oversight. The cycle time, failure modes and intervention rates behind that number matter as much as the success rate. Watch for production deployment timelines with Mercedes-Benz, availability of the model to other robot makers, and safety certification.