Reproduce “LAGOON: Language-Guided Motion Control” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shusheng Xu, Huaijie Wang, Jiaxuan Gao, Yutao Ouyang, Chao Yu, Yi Wu
LAGOON is a framework from Tsinghua University and Shanghai Qi Zhi Institute that translates natural language commands into physically executable robot motions using a combination of human motion generation, motion retargeting, and robust reinforcement learning. It addresses learning from imperfect synthetic demonstrations and demonstrates successful deployment on a real-world quadrupedal robot.
Key insights
- Robust robot policies capable of complex, physically realistic behaviors can be effectively learned from imperfect and physics-violating synthetic motion demonstrations.
- A combined reward strategy, integrating adversarial learning with an optimal-matching-based state-error component, is essential for robustly imitating imperfect, misaligned reference motions.
- Domain randomization during training significantly enhances the sim-to-real transferability and robustness of language-guided robot control policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.