Google DeepMind Unveils Gemini Robotics 2.5 With 96% Factory Success as Speed Rises 30%
Updated
Updated · Ars Technica · Nov 10
Google DeepMind Unveils Gemini Robotics 2.5 With 96% Factory Success as Speed Rises 30%
3 articles · Updated · Ars Technica · Nov 10
Summary
Gemini-powered Apollo 2 robots reached a 96% success rate kitting automotive parts in Mercedes-Benz factories, giving DeepMind one of its clearest real-world tests yet.
Gemini Robotics 2.5 improved lab performance by 30% to 100% over the prior version and ran tasks 20% to 30% faster, with gains aimed at fine-motor work.
More than 200 evaluations included serving food with tongs, using a pipette and tying a bow—tasks that stress dexterity and handling of deformable objects.
DeepMind says the system layers Gemini Embodied Reasoning for live video-based planning over a vision-language-action model that turns instructions into robot movements.
Mercedes-level reliability remains a hurdle: DeepMind says closing the remaining 4% gap with near-error-free human performance will be a steep road, even as it readies an open safety benchmark.
Despite a 96% factory success rate, why does Google's advanced AI robot still fail 64% of the time when screwing in a lightbulb?
As AI robots enter factories, how will Google's new safety benchmarks handle split-second decisions between damaging industrial products and risking human injury?
With Apollo 2 acting merely as a data vacuum, what groundbreaking capabilities are being secretly reserved for the commercial Apollo 3 launch?