Event

Study: Robot policy code can be iteratively learned and transferred across tasks

1 reports · 1 sources

So Kuroki, Kento Kawaharazuka, and Yujin Tang published research that treats complete closed-loop robot policy implementations as reusable execution experience: a coding agent generates policy code based on a small number of successful demonstrations, then iteratively improves it using simulation feedback. On four source tasks in RoboCasa, the success rate rose from 28.3% for initially generated code to 64.2% for code selected via a validation set. On nine other target tasks in RoboCasa, the success rate was 45.2% without reference, 41.5% when using initially generated code from the source tasks, and 57.0% when using code selected via the validation set. The research states that this learning approach was also validated on a real Unitree G1.