Originally published on The Daily Flare.
Google DeepMind introduced Gemini Robotics 2.5 on November 10, 2026, saying the model helps robots learn more dexterous tasks faster and operate in less predictable real-world settings.
The company describes the release as its most capable vision-language-action model for robots so far. It evaluated the system on more than 200 tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid. DeepMind says the model achieved success rates 30% to 100% higher than previous baselines, depending on the task, while completing evaluations 20% to 30% faster.
Some results involved tasks that earlier models had struggled to complete. In figures attributed to DeepMind’s benchmark results, the task of removing a disk drive from its bay rose from a 0% success rate to 100%, as did placing a lime on a top shelf and packing pastries with tongs. Taking a Polaroid photo rose from 0% to 92%, while untying a ribbon bow reached 70%, up from 0%. The results varied: some of the harder multi-step tasks remained between 40% and 75% success.
The tasks test combinations of skills rather than simple gripping, including tool use and manipulating flexible or delicate objects. DeepMind says the new model improved task precision, robustness and speed. The model can also run on different robot designs: Apptronik’s Apollo 2 humanoid has a five-finger SharpaWave hand with 22 degrees of freedom, while the Franka Duo platform uses two-finger parallel grippers.
Factory test in Germany and Hungary
DeepMind, Apptronik and Mercedes-Benz tested the system in a proof of concept on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apptronik’s Apollo 2 robots performed automotive kitting, picking and packing parts with a reported 96% success rate.
The work involved sorting parts from bulk bins into trays in the sequence needed for vehicles on the line. DeepMind said the repetitive task could be handled by robots, leaving technicians free to focus on more complex build work. The 96% figure describes the proof-of-concept performance, not a claim that the robots achieved perfect reliability.
The factory trial used two AI systems for different parts of the job. Gemini Robotics ER, the embodied reasoning model, tracked the workflow and what remained to be done. Gemini Robotics 2.5 controlled the robot’s arms and hands. DeepMind says the robots could reason through multiple steps and adapt their actions to changes in the work environment.
The company’s robotics head, Carolina Parada, told Ars Technica that there can be a trade-off between tailoring a robot to one application and developing a system that generalizes across tasks. “We can be very good at a particular application at the cost of generalization,” she said. “Or we can just push on the generalization axis and maybe not get to 99.9 in every application.”
Simulation and safety work
DeepMind also credited Gemini 4 Argon with helping accelerate development. The company says it used Argon as a simulation engine, generating thousands of variations on practice runs from real-world camera data. The resulting digital twins allowed the team to test tasks in simulation and help the robotics model generalize to new tasks and settings.
Alongside dexterity and reliability, DeepMind said it had improved the model’s physical safety steerability and introduced a benchmark for testing robotics-specific jailbreak attempts. The separate ASIMOV-Jailbreak benchmark measures resistance to disguised unsafe requests. According to IoT Tech News, Google said the embodied reasoning system achieved a 99.5% resistance rate on that benchmark.
Gemini Robotics 2.5 is rolling out to DeepMind’s robotics partners, the company said.
Top comments (0)