Google DeepMind releases Gemini Robotics 2.5, with 96% success in Mercedes-Benz parts picking pilot
The VLA model posts 30% to 100% higher task success than baselines and runs 20% to 30% faster, DeepMind says. It is rolling out to robotics partners.
What happened
Google DeepMind has released Gemini Robotics 2.5. It is a vision-language-action model, which means an AI that looks at a scene, understands instructions and then controls a robot's movements. The company says it helps robots handle tools and delicate objects.
Why it matters to you
Robots that cope with messy real workplaces could take over repetitive prep work in factories and leave skilled tasks to human workers.
The takeaway
Google says its new robot AI did a factory parts-sorting job correctly 96% of the time in a pilot.
What it means for you
DeepMind says the model works natively across robot types, including Apollo 2 with the SharpaWave hand and the Franka Duo, and it released the open-source ASIMOV-Jailbreak benchmark.
Manufacturers can look at the Mercedes-Benz pilot, where Apollo 2 sorted parts from bulk bins into trays, as an example of the kitting work DeepMind targets.
DeepMind reports a 96% pilot success rate and a rollout to partners including Agile, Apptronik and Mercedes-Benz, but gives no commercial terms or revenue figures.