A breakthrough in the field of robotics.

The robotic arm learns to brew coffee automatically using its brain. Photo: Physical Intelligence .

The field of robotics has just witnessed a significant step forward as Physical Intelligence, a San Francisco-based technology startup, announced a new artificial intelligence model called "π0.7".

The biggest difference of this model lies in its ability to "generalize by combination." This means that the robot is no longer just mechanically repeating what it is taught, but can now reason independently to solve novel situations.

In real-world tests, researchers witnessed surprising results. Most notably, a robotic arm that had never been trained to use an air fryer was able to independently bake sweet potatoes.

This robot independently figured out how to open the lid, place food inside, and operate the device accurately. This capability is similar to how large language models (LLMs) like GPT-4 can process complex questions based on learned knowledge.

Physical Intelligence aims to raise $1 billion with a valuation of $11 billion . Photo: Physical Intelligence.

"It was a surprising moment for us to see how the model could combine skills in that way," said Lucy Shi, an AI researcher at Physical Intelligence.

She emphasized that the ability to spontaneously develop new behaviors without specific sample data is key to robots adapting to a constantly changing real-world environment.

Another significant advantage of π0.7 is its flexibility across different hardware platforms. The model can effectively apply knowledge from one type of robotic arm to another with a different physical structure. The robot can even automatically adjust its grip angle to suit the characteristics of the new device without human intervention.

"We found that the model can apply strategies learned from one robot to another robot with a completely different shape," Lucy Shi further shared about the ability to share knowledge between different machine lines.

In terms of performance, the new model has achieved impressive figures. In certain tasks, π0.7 achieved a success rate of approximately 85.6%. This figure is very close to the 90.9% achieved by professional robot operators with hundreds of hours of experience. This achievement opens up prospects for the faster integration of multi-purpose robots into daily life.

Training is now simpler as users can adjust the robot's behavior through natural language commands instead of having to load thousands of hours of video data as before.

This breakthrough lies not only in technology but also in attracting significant interest from global investors. With backing from Jeff Bezos, OpenAI, and Thrive Capital, Physical Intelligence is now valued at $2 billion .

The company aims to create a unified "brain" capable of controlling all types of robots to perform any physical task in the real world . This is considered a significant breakthrough, moving artificial intelligence beyond computer screens to truly impact and assist humans in complex tasks.