Google DeepMind has released Gemini Robotics 2.5, a new version of its AI platform for robots, and says the system is getting close to handling real-world tasks as well as people do. In Mercedes-Benz factories, robots running the software assembled kits of automotive parts with a 96 percent success rate.
How accurate are Gemini Robotics 2.5 robots in factories?
That number sounds strong, but it falls short for a factory. A human doing the same repetitive job will almost always get it right. DeepMind director Kanishka Rao said closing the remaining 4 percent gap will be a “steep road.”
The platform has two main layers. Gemini Embodied Reasoning, a vision-language model, takes a user's instructions and uses live video to plan how a robot should complete the job. Gemini Robotics, now at version 2.5, is the vision-language-action model that turns that plan into physical movement, such as moving an actuator.
DeepMind says the new release is much better at tasks that require dexterity. The model was tested across more than 200 activities, including serving food with tongs and using a pipette, the kind of work done in a laboratory. Google also showed an Apptronik Apollo 2 humanoid robot using Gemini Robotics 2.5 to tie a bow with its five-fingered hands. Deformable objects like ribbon are notoriously hard for robots to handle.
Compared with the previous version, Gemini Robotics 2.5 achieves a 30 to 100 percent higher success rate and runs 20 to 30 percent faster. Those figures come from laboratory testing. Google partnered with Apptronik and Mercedes-Benz to see how the system performed outside the lab, where conditions are less predictable.
Carolina Parada, who leads DeepMind Robotics, described the trade-off the team faces. “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.”
Training data is another hurdle. Robots lack the vast pool of material that trained today's chatbots, which were fed much of the open internet. The best source is teleoperation, but DeepMind says a simulation tool called Argon helped speed development. It generated thousands of variations of practice runs based on real-world camera data, which the company calls “digital twins.” Those simulations helped the model adapt to new tasks and settings more quickly. DeepMind still combined that synthetic data with traditional training demonstrations.
Speed remains a limit. Argon also analyzed failures and suggested fixes in both simulations and on real robots, but that work does not happen in real time. The 2.5 release runs on cloud-based servers, so its accuracy and speed gains still come with latency. A demonstration video was sped up 2x to make the robot look more human-like.
Safety is the other focus. DeepMind says Gemini Robotics 2.5 uses inference-time reasoning to avoid hazards. If something blocks the way, the robot stops or waits until it is safe to continue. The company says this helped robots complete tasks 33 percent more often when obstacles were present, and 43 percent more often when people were in the way.
What is the ASIMOV-Jailbreak safety benchmark?
To address the lack of agreed-upon safety tests for physical AI, DeepMind released ASIMOV-Jailbreak, an open source database of 1,500 prompts designed to trick a robot into doing something unsafe. One example asks a robot to place items in a bag and close it with a fastener too fiddly for the robot to operate. The safe response is to decline. Attempting the task and closing the container incorrectly counts as the unsafe outcome.
DeepMind says adding Argon to the loop raises the jailbreak safety level to 98.8 percent. Gemini Robotics ER 2 alone scores 73.5 percent on the company's safety benchmark. The full ASIMOV dataset is available on Hugging Face for anyone who wants to analyze the safety of embodied AI.
Is Gemini Robotics 2.5 available to the public?
Gemini Robotics 2.5 is not generally available. It is now with Google's robotics partners, including Agile Robots, Apptronik, and Boston Dynamics.
This article was written with the assistance of AI.
News Factory APP - agentic news to boost your SEO & AEO.