Gemini Robotics 2.5: DeepMind's Robot AI Hits 96% in Factories

A 96 percent rate is still short of what a manufacturer needs, since a human performs such repetitive work almost flawlessly. DeepMind director Kanishka Rao called closing the last 4 percent a "steep road." Robotics head Carolina Parada pointed to the trade-off: the team can excel at one application at the cost of generalization, or push generalization and perhaps not reach 99.9 percent everywhere.

To overcome scarce physical training data, DeepMind used Gemini 4 Argon as a simulation engine, generating thousands of variations of practice runs from real camera data. This synthetic data was combined with traditional demonstrations. Argon also analyzed failures. The model still runs on cloud servers, so latency remains.

Safety benchmark

Gemini Robotics 2.5 uses inference-time reasoning to avoid hazards, stopping or waiting when something is in the way. Tasks were completed safely 33 percent more often with obstacles, and 43 percent more when people were in the way.

DeepMind also released ASIMOV-Jailbreak, an open source set of 1,500 prompts designed to trick robot AI into unsafe actions. With Argon in the loop the system reached a 98.8 percent jailbreak safety level, versus 73.5 percent with Gemini Robotics ER 2 alone. The dataset is on Hugging Face. Gemini Robotics 2.5 is not generally available but is with partners including Agile Robots, Apptronik and Boston Dynamics.