The Pulse
Google says Gemini Robotics 2.5 raises task success 30% to 100% over earlier baselines
Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model it says performs more tasks and learns skills faster. The company reported higher task success in evaluations and a 96% success rate in a factory proof of concep
AI.info Team · Written with AI · fact-checked ·
A broader set of robot skills
Google DeepMind introduced Gemini Robotics 2.5 on November 10, 2026, describing it as its most capable vision-language-action model to date. The company says the model expands the range of skills robots can learn and the speed at which they acquire them, with the aim of supporting work in less structured real-world settings.
Google DeepMind says it tested the model’s task breadth, performance and complexity across more than 200 distinct tasks, including serving food with tongs, tying a bow and pipetting liquid. In those evaluations, Gemini Robotics 2.5 achieved success rates 30% to 100% higher than previous baselines, depending on the task, and completed tasks 20% to 30% faster, the company said. The model works with different robot types, according to Google DeepMind, including Apptronik’s Apollo 2 humanoid and the Franka Duo platform.
Factory-floor proof of concept
Google DeepMind partnered with Apptronik and Mercedes-Benz to test Apollo 2 on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. In the proof of concept, Apollo 2 autonomously sorted parts from bulk bins into trays in the sequence required for each vehicle on the line.
The robot used Gemini Robotics ER to track the workflow and Gemini Robotics 2.5 to control its arms and hands, picking up metal brackets and placing them in containers while correcting for gaps. Google DeepMind reported a 96% success rate for the pilot and described it as progress toward wider industrial use. Apptronik co-founder and CEO Jeff Cardenas said the models improved dramatically in a few weeks of post-training and carried over to Mercedes-Benz production with very little drop-off.
Safety testing and steerability
Google DeepMind says Gemini Robotics 2.5 uses guidance during inference to steer around physical hazards, raising safe task completion by 33% around obstacles and 43% near people in its evaluations. The company is also releasing ASIMOV-Jailbreak, an open-source benchmark with hundreds of prompts disguising unsafe requests across dozens of scenes.
Gemini 4 Argon achieved 99.5% jailbreak resistance in the company’s evaluation, compared with 82% for its previous baseline, Gemini Robotics ER 2. The company says Argon leads its agentic robotics evaluations across unsafe-task refusal, resolving uncertainty and working near people.
Argon’s role in development
Google DeepMind says Gemini 4 Argon helped develop Gemini Robotics 2.5 by serving as a simulation engine for robotics research. Argon can generate thousands of simulated practice runs from a real-world camera scene and build a digital twin, the company said.
Google DeepMind says simulated runs can be combined with real training demonstrations to help robots generalize to new scenes and objects. The company also says Argon can help researchers investigate task failures from video and test potential improvements in simulation and on real robots.