Reproduce “MonoDream: Monocular Vision-Language Navigation with Panoramic Dreaming” 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
Shuo Wang, Yongcai Wang, Zhaoxin Fan, Yucheng Wang, Maiyue Chen, Kaihui Wang, Zhizhong Su, Wanting Li and 3 more
MonoDream is a framework enabling monocular Vision-Language Navigation (VLN) agents to achieve performance comparable to panoramic RGB-D methods. It leverages a Unified Navigation Representation (UNR) supervised by Latent Panoramic Dreaming (LPD) tasks to implicitly model global, geometric, and future environmental contexts from limited monocular observations, then operates efficiently during inference with only monocular input, achieving state-of-the-art results on RxR-CE Val-Unseen (SR 49.4%, SPL 40.9%).
Key insights
- Monocular VLN agents can develop an internal, holistic understanding of their environment, including global, geometric, and future contexts, without explicitly reconstructing panoramic or depth maps during inference.
- Supervising a unified latent representation with 'latent panoramic dreaming' for both current and future panoramic RGB and depth features effectively enriches the agent's spatial and temporal awareness from constrained monocular inputs.
- Aligning diverse navigation-relevant visual semantics with language-grounded action intent within a shared Unified Navigation Representation (UNR) enhances action prediction and fosters robust cross-dataset generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.