Reproduce “Hi-WM: Human-in-the-World-Model for Scalable Robot Post-Training” 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
Yaxuan Li, Zhongyi Zhou, Yefei Chen, Yanjiang Guo, Jiaming Liu, Shanghang Zhang, Jianyu Chen, Yichen Zhu
The Hi-WM framework introduces a scalable robot post-training paradigm by shifting human-in-the-loop intervention from physical robots to an interactive, action-conditioned world model. It improved real-world task success rates by an average of 37.9 percentage points over base policies while demonstrating substantial cost savings of up to nearly $100,000 USD at scale.
Key insights
- Virtualizing human-in-the-loop robot training through a high-fidelity world model significantly reduces operational costs and enhances the scalability of data collection.
- Targeted human corrections around failure states, facilitated by state rollback and trajectory branching, provide denser and more effective supervision for policy improvement.
- Training the world model with high-dimensional action conditioning, failure data, and edge-case coverage is critical for ensuring its fidelity and the transferability of virtual interventions to the real world.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.