🤖 Gemini Robotics 2.5: Trials at Mercedes-Benz Plants
On November 10, 2026, DeepMind unveiled its most advanced VLA model: it controls a robot's arms and grippers based on camera frames and text commands. Across 200+ tasks, success rates increased by 30–100%, and speed improved by 20–30%, while the Apptronik Apollo 2 humanoid performed kitting with 96% success on Mercedes-Benz assembly lines.
🌍 VLA models are taking over routine kitting at auto plants, and the combination of reasoning and motor skills serves as a control template for robotics and OEMs. The ASIMOV-Jailbreak benchmark makes safety measurable: metrics in proximity to humans increased by 43%.
👤 Early access to Gemini Robotics 2.5 is available via a waitlist, with the blog and model page accessible, and the google/asimov_agentic dataset on Hugging Face can be downloaded and tested locally. A 96% success rate on an assembly line is a rare verifiable metric in robotics.
Source 1: https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/