Reproduce “HarmoWAM: Harmonizing Generalizable and Precise Manipulation via Adaptive World Action Models” 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
Qiuxuan Feng, Jiale Yu, Jiaming Liu, Yueru Jia, Zhuangzhe Wu, Hao Chen, Zezhong Qian, Shuo Gu and 3 more
HarmoWAM introduces a unified World Action Model that resolves the trade-off between generalizable robot movement and precise manipulation by adaptively coordinating a predictive expert and a reactive expert. The framework achieved an 89% average success rate in-domain and maintained an 81% success rate across out-of-domain scenarios with only a 7.9% performance drop.
Key insights
- A systematic empirical study revealed a fundamental trade-off in existing WAMs between generalizable transit and precise manipulation, motivating the need for a unified approach.
- Dynamically coordinating complementary action experts, a predictive one for precision and a reactive one for generalization, allows a single system to address both aspects.
- Incorporating latent features from a generative world model into both action experts significantly enhances their ability to model physical dynamics and maintain temporal coherence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.