W-RIZZ: A Weakly-Supervised Framework for Relative Traversability Estimation in Mobile Robotics

TL;DR

W-RIZZ tackles traversability estimation for mobile robots in unstructured environments by using sparse relative annotations and cross-image labeling within a mean-teacher framework. It introduces a squared-hinge based loss for relative traversability (L_RIZZ) and a consistency term, enabling accurate per-pixel traversability scores in [0,1] without dense pixel labels. The approach demonstrates strong offline performance on the WayFAST dataset and robust real-world navigation on a TerraSentia robot, outperforming a state-of-the-art self-supervised baseline and approaching strongly supervised segmentation with far fewer annotations. The results highlight improved cross-image consistency, better domain generalization, and practical practicality, making the method appealing for scalable deployment in field robotics, with potential extensions to pseudo-labeling and foundation-model integration.

Abstract

Successful deployment of mobile robots in unstructured domains requires an understanding of the environment and terrain to avoid hazardous areas, getting stuck, and colliding with obstacles. Traversability estimation--which predicts where in the environment a robot can travel--is one prominent approach that tackles this problem. Existing geometric methods may ignore important semantic considerations, while semantic segmentation approaches involve a tedious labeling process. Recent self-supervised methods reduce labeling tedium, but require additional data or models and tend to struggle to explicitly label untraversable areas. To address these limitations, we introduce a weakly-supervised method for relative traversability estimation. Our method involves manually annotating the relative traversability of a small number of point pairs, which significantly reduces labeling effort compared to traditional segmentation-based methods and avoids the limitations of self-supervised methods. We further improve the performance of our method through a novel cross-image labeling strategy and loss function. We demonstrate the viability and performance of our method through deployment on a mobile robot in outdoor environments.

Figures (7)

• Figure 1: Our proposed framework trains a model for traversability estimation using sparse pairwise labels of relative traversability. The camera images are from the dataset introduced by Gasparino et al.wayfast. In the traversability map, warmer colors indicate higher anticipated ease of traversal.
• Figure 2: Illustration of our annotation strategy on the WayFAST dataset wayfast, showing intra-image labeling (top) and cross-image labeling (bottom). Yellow crosshairs indicate both points in a pair are equally traversable, while the blue point is more traversable in pairs with blue and red crosshairs.
• Figure 3: Validation set $\text{HDR}_\tau$ at various thresholds as a function of the number of annotated training images.
• Figure 4: Example inference visualizations from our validation split of the WayFAST dataset wayfast. The input color images are shown in the top row, and the corresponding traversability predictions are shown in the bottom row. The traversability score ranges from 0 to 1 for each image, with a higher traversability score indicating that a region is more easily traversed.
• Figure 5: Predictions on images from the WayFAST dataset wayfast, showing color images (top), as well as predictions from a model trained without cross-image labeling (middle) and with cross-image labeling (bottom).