Reproduce “V-STRONG: Visual Self-Supervised Traversability Learning for Off-road Navigation” 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
Sanghun Jung, JoonHo Lee, Xiangyun Meng, Byron Boots, Alexander Lambert
V-STRONG introduces a self-supervised learning framework that integrates a Vision Foundation Model (SAM) with vehicle trajectory data to predict terrain traversability in complex off-road environments. This method achieves an AUROC of 0.959 on the RELLIS-3D dataset and demonstrates improved generalization to unseen terrains, resulting in lower collision rates during autonomous navigation.
Key insights
- Vision Foundation Models (VFMs) like SAM offer robust, generalized visual features and fine-grained instance segmentation, which are effective priors for self-supervised traversability learning.
- Integrating SAM's mask proposals with vehicle trajectory data allows self-supervised methods to overcome overfitting to narrow paths, enabling the identification of broader traversable regions.
- A pixel-level contrastive learning framework, combining both trajectory and mask-derived signals, leads to a traversability predictor with enhanced generalization capabilities in diverse off-road settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.