SyNeT: Synthetic Negatives for Traversability Learning

TL;DR

SyNeT tackles the lack of explicit negative examples in self-supervised traversability learning by synthesizing scene-coherent negative regions with a diffusion-based pipeline and integrating them into PU and PN training. The method couples a positive label source from robot trajectories with synthetic negatives to form stable decision boundaries, reinforced by a novel negative-center and repulsion losses. An object-centric FPR evaluation provides a labeling-free metric to quantify performance on inserted negatives, and extensive experiments show improved robustness and generalization across indoor, off-road, urban, and social settings. The approach is demonstrated on multiple datasets, with practical on-robot deployment and publicly released code, underscoring its potential for safer autonomous navigation in diverse environments.

Abstract

Reliable traversability estimation is crucial for autonomous robots to navigate complex outdoor environments safely. Existing self-supervised learning frameworks primarily rely on positive and unlabeled data; however, the lack of explicit negative data remains a critical limitation, hindering the model's ability to accurately identify diverse non-traversable regions. To address this issue, we introduce a method to explicitly construct synthetic negatives, representing plausible but non-traversable, and integrate them into vision-based traversability learning. Our approach is formulated as a training strategy that can be seamlessly integrated into both Positive-Unlabeled (PU) and Positive-Negative (PN) frameworks without modifying inference architectures. Complementing standard pixel-wise metrics, we introduce an object-centric FPR evaluation approach that analyzes predictions in regions where synthetic negatives are inserted. This evaluation provides an indirect measure of the model's ability to consistently identify non-traversable regions without additional manual labeling. Extensive experiments on both public and self-collected datasets demonstrate that our approach significantly enhances robustness and generalization across diverse environments. The source code and demonstration videos are publicly available at the project page: https://anonymous-synet.github.io/SyNet.github.io/

Figures (9)

• Figure 1: High-level concept of our synthetic negative-based traversability estimation. (Top) The robot treats its driving path as positive and introduces synthetic objects as explicit negatives. (Bottom) The pipeline takes an input image, generates a synthetic negative, and produces a similarity map for traversable regions.
• Figure 1: Overview of the datasets used for training and quantitative evaluation in our experiments
• Figure 2: Synthetic negative generation pipeline. (1) Region selection: randomly sample a region of interest (ROI) and target object size within the predefined ground area. (2) Inpainting: generate synthetic negatives using the Stable Diffusion 3.5 stabilityai_sd35_2024 inpainting pipeline and FLUX.1 Fill bfl_flux1_tools_2024. (3) Segmentation & filtering: segment the generated object with LangSAM medeiros_lang_segment_anything_2023 and apply scene filters (e.g., object count and pixel-area thresholds); if unmet, repeat steps (2) and (3) for the same ROI. (4) Composition: blend the approved negative object into the base image to obtain the final composite and its negative mask.
• Figure 3: Overview of the SyNeT training strategy integrated into LORT (PU) and VS (PN) baselines. The Synthetic Negative Generation Module augments the input image with explicit negatives, allowing the model to extract synthetic negative features ($\widetilde{N}$). These features are then incorporated into the respective learning objectives ($\mathcal{L}_\mathrm{neg}$ for LORT, $\mathcal{L}_\mathrm{contra}$ for VS) to enforce a clear separation between traversable and non-traversable regions in the feature space.
• Figure 5: Overview of the RUOS dataset. The driving trajectories are visualized across three distinct scenarios: Campus, Convention Center, and Public Park. Each color corresponds to a recording sequence, highlighting the diversity of the collected paths.