Gemini Robotics is moving beyond lab demonstrations. On Nov 10, 2026, DeepMind revealed Gemini Robotics 2.5, a new generation of its AI-powered robotic platform, alongside an open safety benchmark called ASIMOV-jailbreak. Google says the system is getting tantalizingly close to completing real-world tasks as well as humans do.
The upgrade targets the part of robotics that tends to expose optimistic presentations: dexterity. Gemini Robotics 2.5 was evaluated across more than 200 activities, including serving food with tongs and using a pipette, the kind of task found in a laboratory. Google also showed Apptronik’s Apollo 2 humanoid robot tying a bow with its five-fingered hands — an awkward little triumph, because deformable objects such as ribbon are notoriously difficult for robots to manipulate.
Across those evaluations, Gemini Robotics 2.5 achieved a 30 to 100 percent higher success rate than the last version and ran 20 to 30 percent faster. At some tasks, it was up to twice as successful. Robots, it turns out, still struggle with ribbon, but at least they are now struggling with improved metrics.
From robotic demonstrations to factory work
Google partnered with Apptronik and Mercedes-Benz to test the system against the uncertainty outside robotics labs. Gemini-powered Apollo 2 robots kitted automotive parts in a few European factories and reached a 96 percent success rate, giving the platform a more demanding proving ground than a carefully arranged demonstration.
That result leaves a 4 percent gap, and DeepMind director Kanishka Rao calls closing it a “steep road.” DeepMind Robotics head Carolina Parada described the central trade-off: “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.”
The platform has several computational layers between a user’s instruction and the robot’s movement. Gemini Embodied Reasoning, the vision-language model component, takes instructions and live video, then plans how the robot can complete the task. A separate vision-language-action, or VLA, component generates the robot’s actions.
Training those layers requires useful examples, and the best training data for robots comes from teleoperation. The team says Gemini 4 Argon helped create a “simulation engine” that accelerated development of Gemini Robotics 2.5 by generating thousands of variations on practice runs based on real-world camera data.
Those variations became “digital twins” of the practice runs, allowing Gemini Robotics 2.5 to generalize new tasks and settings more quickly. DeepMind combined the synthetic data with traditional training demonstrations to ground the model, then used Argon to analyze failures and suggest improvements in simulations and real robots.
Safety still has to survive contact with reality
Argon does not work in real time, and the new VLA is not lightning quick either. Google has experimented with embedded robotics models, but the current 2.5 release still runs on cloud-based servers — a practical limitation for robots expected to respond around people.
DeepMind is also working to ensure failures do not put nearby humans at risk. Gemini Robotics 2.5 uses inference-time reasoning to avoid hazards: if something gets in the way, the robot stops completely or waits until it is safe to continue.
According to DeepMind, that feature helped Gemini robots complete tasks safely 33 percent more often when obstacles were present. The success rate rose 43 percent when people were in the way, which is a useful reminder that the most important object in a robot’s workspace is often the one that was never included in the test plan.
ASIMOV-jailbreak is intended to support the next stage of that work as an open safety benchmark. DeepMind is preparing for robots to operate alongside humans, while Gemini Robotics 2.5 tries to improve task success, speed, generalization and hazard avoidance at the same time.
Based on
- DeepMind reveals Gemini Robotics 2.5 and open safety benchmark “ASIMOV-jailbreak” — arstechnica.com
- DeepMind reveals Gemini Robotics 2.5 and open safety benchmark “ASIMOV-jailbreak” | Ars OpenForum — arstechnica.com
- Are We Ready for Robots That Code Their Own Skills? — spectrum.ieee.org
- Google DeepMind uses AI to prove Signal’s encryption works — axios.com
- Humanoid robots are getting more lifelike — cnbc.com
