SIAgentNews.com
← All news
DevelopingGeminiSecurity

Google's Gemini Robotics 2.5 hits 96% success rate in factory trials

The new vision-language-action model allows humanoid robots to perform complex sorting and assembly tasks with higher speed and safety.

· First reported by interestingengineering.com

96%success rate in factory trials

The short version

  • Gemini Robotics 2.5 achieved a 96 percent success rate in a factory proof-of-concept with Apptronik's Apollo 2 humanoid robots.
  • The model was tested on over 200 tasks, including tying bows, serving food, and handling flexible materials.
  • The system improved safe task completion by 43 percent when operating near humans.
  • Google released ASIMOV-Jailbreak, an open-source benchmark to test if robots can resist disguised unsafe instructions.

Why it matters

This progress suggests that robots are moving closer to handling the unpredictable nature of human environments. It marks a shift toward robots that can learn from real-world footage rather than just being programmed for single actions.

Testing robots on the assembly line

Google DeepMind recently introduced Gemini Robotics 2.5 to help robots learn physical tasks more quickly. The model is designed to help machines handle delicate objects and operate reliably in real-world settings. It aims to bridge the gap between digital instructions and physical movement.

The model was recently put to the test in a factory proof-of-concept involving Apptronik’s Apollo 2 humanoid robots. These robots worked on Mercedes-Benz production lines in Hungary and Germany. They were tasked with sorting automotive parts from bulk bins into organized trays for vehicle assembly.

During these trials, the robots had to autonomously arrange components in the specific sequence required for manufacturing. The system was able to adjust its movements as needed to complete the work. As interestingengineering.com reported, the model achieved a 96 percent success rate during this specific factory evaluation.

How the VLA model functions

Gemini Robotics 2.5 is a vision-language-action model, or VLA. This type of model translates visual information and written instructions into actual physical movements for a robot. It allows the machine to understand what it sees and then decide how to move its limbs to interact with that object.

To develop this capability, Google evaluated the model across more than 200 distinct tasks. These included varied activities like serving food with tongs, tying bows on gift boxes, and transferring liquids using pipettes. These tests were intended to measure the robot's ability to handle flexible materials and use everyday tools.

Compared to previous baselines, the new model showed significant improvements in performance. Google noted that success rates were 30 percent to 100 percent higher depending on the specific task. Additionally, the robots were able to complete these activities 20 percent to 30 percent faster than before.

Safety and navigation improvements

Beyond just moving objects, Google focused on making the robots safer to use around people. The model uses inference-time reasoning guidance to help robots identify and avoid physical hazards. This allows the machines to navigate complex environments without colliding with obstacles or humans.

In company evaluations, these safety updates led to a 33 percent increase in safe task completion around obstacles. Even more notably, the capability to operate safely near people increased by 43 percent. These improvements are critical for moving robots out of isolated cages and into shared workspaces.

The company also introduced a new open-source benchmark called ASIMOV-Jailbreak. This tool is designed to test whether robotic systems can resist disguised unsafe instructions. Google reported that their Gemini 4 Argon model achieved 99.5 percent resistance to these types of jailbreak attempts.

Simulated training and digital twins

A major part of the development process involves using a model called Gemini 4 Argon. This system helps accelerate the creation of new robotics models by generating thousands of simulated practice runs. It creates digital twins, which are virtual copies of physical environments, based on real-world camera footage.

Researchers can combine these digital simulations with real training demonstrations to teach robots new skills. This method helps robots generalize what they have learned to unfamiliar objects or different settings. It allows the machine to practice in a risk-free virtual space before trying the action in the real world.

Argon also provides a feedback loop by analyzing failed attempts during training. It can then suggest specific improvements for subsequent training sessions. This iterative process helps the robots refine their movements and increase their reliability over time.

The path to commercial use

While the results are promising, Google clarified that the 96 percent success rate is currently a proof-of-concept. It does not yet represent evidence of widespread commercial deployment across the industry. The technology is still in the phase of being refined and tested in controlled environments.

The company is currently rolling out Gemini Robotics 2.5 to its existing robotics partners. These partners will use the model to improve the dexterity and reliability of their own hardware. The goal is to eventually automate repetitive and physically demanding tasks in industrial settings.

By handling these heavy or boring jobs, the technology aims to free up human technicians for more complex assembly work. This shift could change how factories operate by allowing humans to focus on high-level problem solving. For now, the focus remains on making robots more adaptable to the nuances of the physical world.

What to watch

  • How quickly the model moves from factory proof-of-concepts to widespread commercial deployment.
  • Whether the 43 percent increase in safety near humans holds up in unscripted public environments.

Try it

Look up the 'ASIMOV-Jailbreak' benchmark to see how researchers test for safety in AI systems.

Sources and links

First reported by interestingengineering.com. 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