SIAgentNews.com
← All news

Google DeepMind launches Gemini Robotics 2.5 for better physical intelligence

The new vision-language-action model helps robots perform complex, delicate tasks with higher success rates and improved safety in real-world settings.

· First reported by Google DeepMind

96%Success rate for picking and packing automotive parts

The short version

  • Gemini Robotics 2.5 achieves a 30% to 100% higher success rate on various tasks compared to previous baselines.
  • The model completes tasks 20% to 30% faster than earlier versions.
  • In a pilot with Mercedes-Benz, the Apollo 2 humanoid robot achieved a 96% success rate in automotive kitting.
  • The model includes new safety features that boost safe task completion around obstacles by 33% and nearby people by 43%.
  • Google is releasing ASIMOV-Jailbreak, an open-source benchmark to test robot safety against disguised unsafe requests.

Why it matters

This advancement moves robots out of controlled labs and into unpredictable human environments like factories and homes. By improving dexterity and safety, it makes it more feasible for robots to handle delicate objects and work alongside people without constant supervision.

Advancing robot dexterity

Google DeepMind has introduced Gemini Robotics 2.5, which is their most capable vision-language-action (VLA) model to date. A VLA model is a type of system that allows a robot to process visual information and language instructions to determine the physical actions it needs to take. This new version is designed to help robots learn new skills more quickly and perform them with greater precision in the real world.

The model was tested across over 200 distinct tasks to measure its breadth and complexity. These tasks include everyday actions like serving food with tongs, tying a bow on a gift box, and pipetting a liquid. By mastering these, the model aims to achieve human-level dexterity by combining advanced tool use with the ability to handle delicate or flexible objects.

According to Google DeepMind, the new model shows significant improvements in robustness and speed compared to previous versions. Depending on the specific task, the success rate is 30% to 100% higher than earlier baselines. Additionally, the robots are now able to complete these various tasks 20% to 30% faster than before.

Real-world industrial testing

To prove the model works in active workplaces, Google DeepMind partnered with Apptronik and Mercedes-Benz. They deployed Apptronik’s Apollo 2 humanoid robot on automotive production lines in Hungary and Germany. The goal was to see if the robot could handle the repetitive and physically taxing task of automotive kitting.

In this process, the robot must sort individual parts from bulk bins into organized trays in a specific order for each vehicle. The Apollo 2 robot uses the Gemini Robotics 2.5 actions model to control its arms and hands while using an embodied reasoning model to keep track of the workflow. This allows the robot to reason over multiple steps and adapt to changes in the environment, such as gaps in the supply bins.

The pilot program demonstrated a 96% success rate for the robot in picking and packing metal automotive brackets. By offloading this monotonous prep work to a robot, the company hopes to keep the assembly line moving more efficiently. This shift allows human technicians to focus on more complex build tasks that require human judgment.

Safety and steerability

As robots move into unstructured environments where people are present, physical safety is a primary concern. Gemini Robotics 2.5 introduces inference-time reasoning guidance to help the robot navigate around hazards in real-time. This system works similarly to safety classifiers, which provide a protective layer of oversight for AI responses, to ensure the robot avoids collisions.

The researchers evaluated these safety features and found that the guidance boosted safe task completion around obstacles by 33%. Furthermore, the model showed a 43% improvement in safe task completion when operating around nearby people. These improvements are essential for moving robots out of isolated cages and into shared human spaces.

Google DeepMind is also focusing on safety for long-horizon tasks, which are complex goals that require many sequential actions. They are releasing a new open-source benchmark called ASIMOV-Jailbreak to test how robots handle disguised unsafe requests. Their Gemini 4 Argon model achieved a 99.5% jailbreak resistance, a major increase from the 82% rate of previous models.

Accelerating research

Developing new skills for robots usually requires massive amounts of physical data and efficient modeling algorithms. To solve these bottlenecks, Google DeepMind used Gemini 4 Argon to act as a simulation engine. This tool helps accelerate the development of Gemini Robotics 2.5 by creating a realistic digital twin of real-world camera scenes.

Argon can generate thousands of varied, simulated practice runs from a single scene, which allows the robot to generalize to new objects more quickly. By combining these simulations with real training demonstrations, researchers can train robots faster without needing to perform every action in the physical world. This approach helps the robot learn from a wider variety of experiences in a shorter timeframe.

The simulation engine also allows researchers to hypothesize why specific tasks might be failing by analyzing video footage. They can then test various improvements in the digital twin before applying them to real robots. This cycle of simulation and real-world testing is a key part of the path toward solving physical artificial general intelligence.

What to watch

  • How quickly these models will be integrated into non-automotive manufacturing.
  • The performance of Gemini Robotics 2.5 when handling highly complex, multi-step household chores.
  • The expansion of the ASIMOV-Jailbreak benchmark to include more diverse human-robot interaction scenarios.

Try it

Check out the ASIMOV-Jailbreak benchmark on Google DeepMind's official research pages to see how they test robot safety.

Sources and links

First reported by Google DeepMind. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news