Google DeepMind's Gemini Robotics 2.5 Hits 96% in Mercedes-Benz Factory Trials

On November 11, 2026, Google DeepMind unveiled Gemini Robotics 2.5, its most capable vision-language-action model for physical AI to date. The launch came alongside a live deployment at a Mercedes-Benz manufacturing plant, where the model handled parts-sorting tasks with a reported 96% success rate — a striking benchmark for a technology that just months ago struggled with real-world unpredictability.

What Changed Between Versions

Gemini Robotics 2.5 is not a minor update. DeepMind reports it completes tasks 20–30% faster than its predecessor and succeeds across a broader range of motion types. In tests covering more than 200 distinct skills — from folding laundry and serving pastries to sorting automotive components — success rates ranged from 30% on the most dexterous challenges up to 100% on well-defined manipulation tasks.

The model also includes reasoning-guided hazard avoidance, so when a robot arm encounters an unexpected obstacle or a potentially dangerous configuration, the model's reasoning layer can intervene before contact occurs. This safety layer is new in the 2.5 release and is explicitly designed for factory and service environments where unpredictable human presence is constant.

The Mercedes-Benz Pilot

The factory pilot combined two DeepMind models: Gemini Robotics ER managed high-level workflow orchestration, while Gemini Robotics 2.5 controlled the arms and hands in real time. Together they handled a continuous parts-handling workflow on an actual production line, not a simulated one. The 96% success rate was measured across the pilot's full runtime — not cherry-picked from a subset of runs.

Mercedes-Benz joins a small group of early robotics partners that have direct API access. That group also includes Agile Robots, Apptronik, and Boston Dynamics. There is currently no public API or pricing announced for broader access.

How It Was Built

To speed up development, DeepMind used Gemini 4 Argon — a separate simulation-focused model — to generate synthetic training environments. This let the team scale data collection without requiring a physical robot for every scenario. The technique is becoming standard in physical AI research, but DeepMind's integration of it directly into the Gemini family is notable.

What It Means

The gap between lab demos and production-line deployment has historically been enormous in robotics. A 96% success rate on a live factory floor, using a vision-language model as the primary controller, suggests that gap is narrowing faster than most analysts expected. The limiting factor now is not the model — it's how quickly robotics manufacturers can integrate these APIs into hardware that factories will actually purchase.

DeepMind's decision to roll this out through a closed partner network rather than a public API means the technology stays inside a controlled surface for now. That will change, but not on a timeline DeepMind has announced.