Google DeepMind expands robotics push with Gemini Robotics 2.5, safety benchmark

Traders Union
All News
Business & Wealth
Companies
DeepMind Gemini Robotics 2.5

Highlights

Google DeepMind's Gemini Robotics 2.5 achieved 30% to 100% higher success rates and 20% to 30% faster execution speeds in lab testing versus the prior version.
Gemini-powered Apollo 2 robots performed parts kitting in Mercedes-Benz European factories with a 96% success rate, but human-level consistency remains difficult.
Adding Argon to the safety loop raised jailbreak safety to 98.8%, compared to 73.5% for Gemini Robotics ER 2, with public release to selected robotics partners.

Robotics upgrade targets factory and dexterity gains

As reported by Ars Technica, Google DeepMind says Gemini Robotics 2.5 delivers stronger performance on real-world manipulation tasks, with overall success rates 30% to 100% higher than the previous version and execution speeds 20% to 30% faster in laboratory testing.

The system combines Gemini Embodied Reasoning, which interprets user instructions and live video to plan tasks, with the Gemini Robotics vision-language-action model that generates physical movements. DeepMind says the latest version performs better on high-dexterity activities across more than 200 tasks, including using tongs to serve food and operating a pipette in lab-style work.

Google also demonstrated Apptronik's Apollo 2 humanoid robot tying a bow with its five-fingered hands, a task that highlights progress in handling deformable objects. In trials with Apptronik and Mercedes-Benz, Gemini-powered Apollo 2 robots carried out automotive parts kitting in several European factories and reached a 96% success rate, although DeepMind executives say closing the remaining gap with human-level consistency remains difficult.

DeepMind frames the platform as a general-purpose robotics system rather than a narrowly programmed automation tool. That approach broadens the range of tasks the robots can attempt, but it also creates trade-offs between generalization and the near-perfect accuracy industrial users may require.

Safety testing and deployment implications

To improve training, DeepMind says it uses Gemini 4 Argon to generate large numbers of simulated task variations from real-world camera data. The company combines those synthetic environments with traditional demonstration-based training, using Argon as well to analyze failures and suggest improvements for both simulations and physical robots.

The current release still runs on cloud-based servers rather than fully embedded hardware, which means latency remains a constraint even as the model improves. That could limit responsiveness in some industrial settings where robots need to act without long pauses.

DeepMind says Gemini Robotics 2.5 uses inference-time reasoning to avoid hazards, stopping or waiting when obstacles interfere with a task. According to the company, that lifted safe task completion by 33% when obstacles were present and by 43% when people were in the way.

The company is also releasing ASIMOV-Jailbreak, an open-source set of 1,500 prompts designed to probe whether robotic AI can be tricked into unsafe behavior. DeepMind says adding Argon to the safety loop produces a 98.8% jailbreak safety level, versus 73.5% for a system using Gemini Robotics ER 2 alone. Gemini Robotics 2.5 is not generally available, but DeepMind says it is already with robotics partners including Agile Robots, Apptronik and Boston Dynamics.

Earlier, we discussed Nvidia’s call for international AI safety dialogue with Chinese companies as governments weigh stronger safeguards.

This material may contain third-party opinions, none of the data and information on this webpage constitutes investment advice according to our Disclaimer. While we adhere to strict Editorial Integrity, this post may contain references to products from our partners.