A Chinese AI model can transfer learned skills between different robots—no fresh training required

Robot skills can move across different hardware designs

According to НВ — Техно: Feagine, a Chinese company, has introduced an AI model named Fi0 that lets learned skills be passed from one robot to another with a different body design, without requiring the model to be retrained. Rather than treating each machine as a blank slate, the system is designed to reuse what a robot has already mastered. To test the idea, the company developed three soft manipulators with different specifications: A01, A02, and A03.

• A01: one flexible segment, two degrees of freedom, weighs 750 grams, and can carry loads up to 200 grams.
• A02: two segments and four degrees of freedom, with a payload of 400 grams.
• A03: three segments, 6+1 degrees of freedom, 50 centimeters long, and able to carry up to 600 grams.

Fi0 works by analyzing the task, the robot's structure, and its current state. Feagine says a single human demonstration may be enough for the robot to perform a new task. The model examines objects, action sequences, contact moments, and the final result of the task. The company calls this approach cross-embodiment learning.

The testing robots were soft manipulators because they are more difficult to control than conventional robotic arms. Fi0 relies on an Embodiment Graph to represent the physical structure and condition of a robot. The graph holds data about the machine's shape, sensors, movement mechanisms, and current state, and it also includes components that analyze the environment, interpret human demonstrations, and predict action outcomes.

Where the technology could go next

Feagine's concept does not center on a single universal humanoid robot. Instead, the company envisions multiple robotic platforms drawing skills from a shared AI model. Fi0 is still at an early stage, and its claimed capabilities need to be verified across more robots, tasks, and real-world settings. The stated goal is for skills learned by one robot to keep their value even after the robot is replaced by a differently built machine.

Interesting Engineering covered Feagine's work on August 17 at 1:30 PM. If Fi0 lives up to its promise, it could cut the time and cost involved in getting robots ready for new jobs. The technology may eventually support fields such as manufacturing and healthcare, where adaptive robots could be deployed more quickly instead of undergoing lengthy training cycles.

As advancements in robotics continue to evolve, the development of humanoid robots that adapt to human movements in real time showcases the potential for even greater integration of AI technologies in robotic systems. This trend highlights the importance of building adaptive capabilities in robots, which could complement the cross-embodiment learning approach demonstrated by Feagine's Fi0 model.