Teaching robots new skills is going to be a lot easier with LLMs. SDS is a see it, do it, and sorted approach that involves quadrupeds learning new skills from a single demonstration. It uses GPT-4o to analyze a video and create reward functions that guide the robot in imitating movements like walking, hopping and running. This system learns through reinforcement learning and improves it performance over time by refining the reward functions.
For this project, a Unitree Go1 robot was used to perform various movements with accuracy. SDS works better than previous methods because it doesn’t need large datasets to train the robot.
[HT] [Authors: Jeffrey Li, Maria Stamatopoulou, and Dimitrios Kanoulas]