Reproduce “AR Forcing: Towards Long-Horizon Robot Navigation World Model” 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
Yifei Yang, Zehua Fan, Huan Li, Aoqi Wang, Lida Huang, Haibao Yu, Haiyan Liu, Xuanyao Mao and 4 more
Researchers from Tsinghua University and Lenovo developed AR Forcing, an autoregressive training strategy for diffusion-based robot navigation world models. This method mitigates exposure bias by integrating model-generated predictions into the training context, yielding consistent long-horizon predictions for up to 16 seconds and improving closed-loop trajectory planning performance across various navigation benchmarks.
Key insights
- Explicitly aligning the training context distribution with the inference context distribution directly addresses exposure bias, improving long-horizon prediction stability.
- Training with model-generated contexts allows the diffusion world model to learn robustness to its own prediction errors without requiring complex auxiliary components.
- The standard conditional diffusion loss is sufficient for providing robust action-to-outcome supervision when applied within an autoregressive training framework, outperforming distribution-matching objectives in trajectory planning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.