Google unveils Gemini Robotics 2.5 for factory humanoids
Google DeepMind's new vision-language-action model allows humanoid robots to perform complex industrial tasks with high precision and safety.
The short version
- Gemini Robotics 2.5 handles flexible objects and everyday tools in unstructured environments.
- The model achieved a 96% success rate during a proof-of-concept on Mercedes-Benz production lines.
- The software is hardware-agnostic and runs on both humanoid hands and parallel grippers.
- Google reports a 99.5% jailbreak resistance rate on a new safety benchmark for robots.
Why it matters
This move signals a shift from laboratory demonstrations to practical, large-scale industrial automation. By enabling robots to handle unpredictable tasks, it paves the way for humanoids to take over monotonous factory work.
Moving beyond the lab
Google DeepMind has introduced Gemini Robotics 2.5, a vision-language-action model designed to move robots out of controlled environments. This model is built to help machines manage everyday tools and delicate, flexible objects in settings where tasks are less predictable. It represents a significant step toward robots that can navigate the messiness of the real world.
The company tested the system across an evaluation set of 200 distinct tasks to measure its dexterity and reliability. The goal is to provide robots with whole-body intelligence, allowing them to perform complex movements rather than just repeating a single motion. This release pushes Google deeper into the field of physical AI, specifically for use on factory floors.
By focusing on unstructured settings, the model aims to solve the problem of robots failing when faced with slight changes in their surroundings. Instead of needing a perfectly organized workspace, these machines are being trained to adapt to the variety of objects they encounter. This shift is essential for moving robotics from specialized niches into general industrial use.
Real-world factory testing
The technology was recently deployed on Apptronik’s Apollo 2 humanoid robot during a project with Mercedes-Benz. In plants located in Hungary and Germany, the robots performed automotive kitting on active production lines. This involved sorting various parts from bulk bins into organized trays in the specific order required for each vehicle.
To manage the complexity of these tasks, Google used two different models to split the workload. The Gemini Robotics ER model tracked the overall workflow of the assembly line, while the Gemini Robotics 2.5 actions model specifically controlled the robot's arms and hands. This division of labor allows the robot to understand the
Versatile hardware compatibility
While the model was showcased on the Apollo 2 humanoid, it is not restricted to a single type of robot design. The software is flexible enough to run on different platforms, such as the Franka Duo, which utilizes parallel grippers instead of humanoid hands. This means the technology can be applied to a wide variety of industrial machines.
The Apollo 2 robot features a SharpaWave hand that provides 22 degrees of freedom to mimic human-like movement. However, because Gemini Robotics 2.5 does not require a humanoid hand to function, companies can choose the hardware that best fits their specific factory needs. This versatility allows for broader adoption across different types of manufacturing equipment.
Simulation and safety research
Google credits Gemini 4 Argon with accelerating the underlying research for these robotics models. The system functions as a simulation engine that generates practice runs from camera scenes to build a digital twin of each task. This allows engineers to test and understand why a robot might fail in a virtual environment before they ever apply changes to a physical machine.
Safety is also a primary focus of this research, particularly regarding how robots behave in everyday environments. The new ASIMOV-Jailbreak benchmark was created to measure how well the system resists disguised unsafe requests across various scenes. Google reported that Argon achieved a 99.5 percent jailbreak resistance rate, which is a significant improvement over previous baselines for the Gemini Robotics ER model.
Industry events and connections
The development of these industrial tools fits into a larger ecosystem of connected technology. For example, the Semtech SGP.32 provides a global eSIM for IoT fleets, which helps manage the connectivity of industrial devices. These types of components are essential for the infrastructure that supports modern factory automation and the deployment of smart machines.
Interested readers can learn more about these trends at the IoT Tech Expo, which is scheduled to take place in Amsterdam, California, and London. This event is part of the broader TechEx series and is co-located with other major gatherings, including the AI & Big Data Expo and the Cyber Security Expo. These venues offer opportunities to see how AI and robotics are being integrated into the global economy.
What to watch
- How quickly other manufacturers will adopt humanoid robots for kitting tasks.
- Whether Gemini Robotics 2.5 can be adapted for more delicate tasks like electronics assembly.
Sources and links
- ReportingIoT News: the article we sourced
First reported by IoT News. 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.