DeepMind launches Gemini Robotics 2.5 with 96 per cent success in Mercedes-Benz parts packing

Google DeepMind has announced Gemini Robotics 2.5, which it describes as its most capable vision-language-action model so far. SiliconANGLE reports that it builds on capabilities from Gemini 4 Argon, the frontier model rolled out in late September.

The model was tested across more than 200 distinct tasks. DeepMind points to gains in dexterity, including handling delicate, flexible objects with everyday tools, as well as reliability and safety. Success rates range from 30 per cent to 100 per cent depending on the task, and tasks finish 20 to 30 per cent faster than with the previous version.

In real-world proof-of-concept tests at Mercedes-Benz, the model achieved a 96 per cent success rate packing automotive parts. It works with several robot types, including Apptronik's Apollo 2 humanoid and the Franka Duo platform, and is rolling out through Agile Robots SE, Apptronik and Boston Dynamics. A new jailbreak benchmark tests whether the model can be steered into unsafe physical actions.

Apptronik chief executive Jeff Cardenas said: "This is what we built Apollo for: useful work alongside people in real industrial settings." On the same day DeepMind announced AlphaGenome 2, which predicts how DNA variants affect specific cell types and adds DNA methylation; it is not approved for clinical use.

Why it matters

The wide spread in task-level success rates is a reminder that headline figures hide uneven performance. Factory deployment will depend on the specific task, not the average.

Adding a safety benchmark for physical jailbreaks is a sign that robot foundation models are being judged on misuse as well as capability.

Indicops view

Indian manufacturers considering robot pilots should test on their own tasks and lighting, and ask vendors for task-level success data. Treat vendor trial numbers, such as the Mercedes-Benz result, as indicative rather than guaranteed for a different shop floor.