DeepMind introduces Gemini Robotics 2.5 for robot dexterity
Original title: Gemini Robotics 2.5 accelerates real-world general dexterity
DeepMind reports 30% to 100% higher success, faster tasks, and 96% kitting success on Mercedes-Benz lines.
Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model meant to widen robot skills and speed learning. It was evaluated on more than 200 tasks and works with Apptronik Apollo 2 and Franka Duo robots. A Mercedes-Benz pilot used Apollo 2 for automotive kitting in Hungary and Germany. DeepMind also released the ASIMOV-Jailbreak benchmark and credited Gemini 4 Argon with accelerating the research.
Unconfirmed: Success rates and jailbreak scores are DeepMind's own evaluations, not independently verified.
- The model is rolling out to robotics partners; public weights are not described.
- DeepMind says safe completion rose 33% around obstacles and 43% near people.
- Gemini 4 Argon scored 99.5% on ASIMOV-Jailbreak versus 82% for Robotics ER 2.
- Apollo 2 uses a five-finger SharpaWave hand with 22 degrees of freedom.
- Vendors
- Google DeepMindApptronik
- Products
- Apollo 2Franka DuoSharpaWave
- Organizations
- Mercedes-Benz
From the original article
5 min read at deepmind.google
Our most capable vision-language-action model enhances robot dexterity, reliability, and safety
By Carolina Parada · deepmind.google
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.