SkillMimic-V2: Learning Robust and Generalizable Interaction Skills from Sparse and Noisy Demonstrations
Runyi Yu, Yinhuai Wang, Qihan Zhao, Hok Wai Tsui, Jingbo Wang, Ping Tan, Qifeng Chen
2025-05-06
Summary
This paper talks about SkillMimic-V2, a new method that helps AI learn skills from examples, even when those examples are messy or there aren't very many of them.
What's the problem?
AI usually needs lots of clear and perfect examples to learn new skills, but in real life, the demonstrations it gets can be incomplete or full of mistakes, which makes learning hard.
What's the solution?
The researchers used special ways to create more practice data and smartly pick which examples to learn from, so the AI can still pick up skills and use them in new situations, even if the original demonstrations weren't great.
Why it matters?
This matters because it means AI can learn to do things more like humans do, handling real-world messiness and becoming more useful in everyday situations where perfect training data isn't available.
Abstract
The approach uses data augmentation techniques and adaptive sampling to improve robust skill acquisition and generalization in Reinforcement Learning from Interaction Demonstration despite noisy and sparse demonstrations.