Researchers taught a Unitree G1 humanoid to learn new skills without forgetting old ones
A new neural network architecture allows humanoids to master sequential tasks while reducing training time and resource usage.
Researchers released a paper describing a method to keep humanoid robots learning continuously without losing their previous abilities. In standard machine learning, teaching a robot a new movement often leads it to override or forget what it learned earlier.
The team developed a system called Similarity-guided LoRA-PNN. This architecture uses a neural network that progressively allocates memory for new tasks while utilizing a lightweight technique called low-rank adaptation to reuse knowledge across skills.
Across six sequentially learned skill categories, our similarity-guided LoRA policy attains the best forward transfer and the highest average accuracy.
To manage this process, the system employs a two-level motion-similarity measure. This tool decides which previous skill the robot should use as a foundation for a new movement, ensuring the machine builds on existing experience rather than starting from scratch.
This approach proved efficient during tests involving six different skill categories. The new method saved nearly 95% of the trainable parameters compared to traditional approaches and reduced training time by over 40%.
The controller achieved a 96.13% success rate during sim-to-sim transfer and was successfully deployed on a physical Unitree G1 robot. Have you ever wondered how complex it is to teach a machine to walk, run, and jump without it tripping over its own digital feet?
Despite these gains, the research relies on specific motion-similarity measurements to function. The study did not test this architecture on hardware other than the Unitree G1, so its performance on different robotic platforms remains to be seen.
Source: arXiv cs.RO