SIAgentNews.com
← All news
DevelopingGeminiSecurity

Google launches Gemini Robotics 2.5 to improve robot dexterity

The new vision-language-action model helps humanoid robots perform delicate tasks and factory work with significantly higher success rates.

· First reported by India Today

96%success rate in car part sorting

The short version

  • Gemini Robotics 2.5 achieved a 96% success rate sorting car parts in tests with Mercedes-Benz and Apptronik.
  • The model improves task success rates by 30% to 100% compared to previous versions.
  • Robots using the model can complete tasks 20% to 30% faster than before.
  • Safety evaluations showed a 43% improvement in safe task completion around people.
  • Google released ASIMOV-Jailbreak, an open-source benchmark to test robot resistance to unsafe instructions.

Why it matters

This technology moves robots beyond simple repetitive motions toward handling delicate objects and navigating human spaces. It signals a shift toward robots that can be deployed in complex, real-world environments like factories and homes.

A leap in robot dexterity

Google has introduced Gemini Robotics 2.5, a new model designed to help robots learn skills more quickly and handle delicate objects with greater care. This update follows the release of Robotics 1.5, which provided foundational features like motion transfer and upper-body control. The new version represents a significant jump in how robots interact with the physical world.

The model is categorized as a vision-language-action (VLA) model. This means it can process visual information, understand language instructions, and translate those inputs into physical actions. By combining these capabilities, the AI allows robots to perform a wider range of tasks with more precision and reliability than previous iterations.

Google tested the model across more than 200 different tasks to see how it handled variety. These tests included complex movements like serving food with tongs, tying a bow on a gift box, and pipetting liquid. While these are not typical factory jobs, they demonstrate the model's ability to manage objects that require careful, nuanced movements.

Performance in the factory

To see how the technology works in a practical setting, Google partnered with Apptronik and Mercedes-Benz. They used Apptronik's Apollo 2 humanoid robots to test the model on automotive production lines in Hungary and Germany. The goal was to automate the repetitive work of picking automotive parts from bulk bins and arranging them into specific trays.

The results from these tests were notable for their high level of accuracy. The Apollo 2 robots achieved a 96% success rate when sorting the car parts into the required order. This level of performance helps automate tasks that would otherwise require human technicians to perform manually.

Beyond just accuracy, the model also improved the speed of these industrial operations. Google reported that Gemini Robotics 2.5 can complete tasks 20% to 30% faster than previous models. It also showed a 30% to 100% higher success rate across various different tasks during testing.

Prioritizing human safety

A major focus of the new model is ensuring that robots can operate safely in environments where people are present. The AI is designed to identify hazards and nearby humans, allowing the robot to adjust its movements in real time. This is essential for moving robots out of isolated cages and into shared workspaces.

Google conducted evaluations to measure these safety improvements specifically. The results showed that the model's safety guidance improved safe task completion around obstacles by 33%. Even more importantly, it improved safe task completion around people by 43%.

To further address the risks of AI, Google is also introducing a new benchmark called ASIMOV-Jailbreak. This is an open-source tool designed to test whether robots can resist unsafe instructions that are disguised as ordinary requests. This testing helps ensure that robots remain predictable and safe in unpredictable environments.

The path to humanoid use

The progression from Robotics 1.5 to Gemini Robotics 2.5 shows a clear trajectory for Google's involvement in the robotics space. By focusing on dexterity and reliability, the company is aiming to make robots capable of working alongside humans in everyday environments. The move toward humanoid forms like the Apollo 2 suggests a desire for robots that can navigate human-centric spaces.

The ability to handle both delicate tasks and repetitive factory work makes the technology versatile for different industries. While the current focus includes automotive production, the model's success with tasks like tying bows suggests it could eventually be applied to domestic or service roles. These improvements are intended to make robots more capable of handling a diverse range of physical interactions.

As the technology matures, the focus will likely shift toward how these robots can be integrated into complex workflows. The combination of high success rates, faster completion times, and improved safety metrics provides a foundation for broader deployment. Google's goal is to create robots that can learn and adapt to the nuances of the physical world more effectively.

What to watch

  • How quickly humanoid robots like Apollo 2 are deployed in non-automotive settings.
  • Whether the 96% success rate holds steady as tasks become more complex and unpredictable.

Try it

Look up the ASIMOV-Jailbreak benchmark to see how researchers are testing robot safety.

Sources and links

First reported by India Today. 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