Gemini Robotics 2.5 hits 96% success in Mercedes-Benz factory trial

DeepMind's Gemini Robotics platform splits the work between two models: Gemini Embodied Reasoning, which plans from live video and instructions, and the vision-language-action model Gemini Robotics, now at version 2.5, which generates the actual robot movements. The new version was evaluated across more than 200 activities, including serving food with tongs and using a pipette. A standout demonstration involved Apptronik's Apollo 2 humanoid tying a bow with its five-fingered hands, a task made difficult because deformable objects like ribbon are hard for robots to manipulate.

The jump in performance comes partly from a simulation engine built with Gemini 4 Argon, which generated thousands of variations on practice runs from real-world camera data. These "digital twins" helped the model generalize new tasks and settings more quickly, though DeepMind still combined the synthetic data with traditional training demonstrations. Argon was also used to analyze failures, but not in real time, and the current 2.5 release still runs on cloud-based servers, so accuracy and speed gains come with some latency.

Safety remains a work in progress. Inference-time reasoning lets the robot stop or wait when obstacles or people are in the way, which DeepMind says allowed tasks to be completed 33 percent more often with obstacles present and 43 percent more often when people were nearby. The open source ASIMOV-Jailbreak benchmark, with 1,500 prompts designed to trick a robotic AI into unsafe actions, is DeepMind's attempt to push the industry toward a common safety standard; placing Argon in the loop yields a 98.8 percent jailbreak safety level, compared with 73.5 percent for Gemini Robotics ER 2 alone.