Google DeepMind's robot model ties bows and serves food with tongs
In its own tests, Google DeepMind says its Gemini Robotics 2.5 model helps robots handle soft, delicate and tool-based tasks more often and more quickly. It is rolling out to the lab's robotics partners.
A robot running Google DeepMind's new model can serve food with tongs, tie a bow on a gift box and pipette a liquid, the lab says. DeepMind introduced Gemini Robotics 2.5 in a post dated 10 November 2026.
It is a vision-language-action model. According to a CellCog write-up of the announcement, that means it takes in what a robot sees and what it is told, and turns that into movement.
DeepMind tested the model on more than 200 tasks. In the lab's own evaluations, it succeeded 30% to 100% more often than earlier systems the lab used as comparison points, depending on the type of task, and finished tasks 20% to 30% faster.
The lab's charts compare it with its earlier Gemini Robotics 2 on single tasks. Packing pastries with tongs rose from 5% to 100% success, and untying ribbon bows from 0% to 70%. Some tasks still fail more often than they succeed: folding trousers and packing a rice cooker each reached 40%.
To test reliability, DeepMind worked with Apptronik and Mercedes-Benz to run Apollo 2 humanoid robots on active car production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The robot sorted parts from bulk bins into trays in the exact order each car needs.
A separate reasoning model, Gemini Robotics ER, kept track of the workflow. Gemini Robotics 2.5 controlled the hands and arms to pick metal brackets from bins and place them in containers. DeepMind reports a 96% success rate. CellCog notes that DeepMind does not say how many picks that figure covers.
Jeff Cardenas, Co-founder and chief executive officer of Apptronik, says in DeepMind's post that the models improved dramatically within a few weeks of extra training and carried over to Mercedes-Benz production with very little drop-off.
For safety, DeepMind says the model uses guidance while it runs to steer around physical hazards. In the lab's evaluations, this raised safe task completion by 33% around obstacles and 43% around nearby people.
DeepMind is also releasing ASIMOV-Jailbreak, a free, open-source (anyone can use and copy it) set of hundreds of written instructions that disguise unsafe requests, to check whether robots can be tricked into unsafe actions.
A wider academic survey of robot hands, not specific to this model, by Ji Pei of Xi'an Jiaotong University, China, says completing a task does not by itself show a robot is ready for deployment. It adds that evidence on long-duration reliability remains limited.
CellCog also reports that the post gives no price, no public access and no technical paper. Gemini Robotics 2.5 is rolling out to DeepMind's robotics partners, the lab says.
Sources
- Gemini robotics 2 5 accelerates real world physical intelligence · deepmind.google
- Gemini robotics 2 5 · cellcog.ai
- Gemini robotics er · docs.cloud.google.com
- arxiv.org · arxiv.org