SyNeT: Synthetic Negatives for Traversability Learning

Anonymous Authors
RA-L (under review)
Paper (PDF) Code Video
Self-Supervised traversability Generation Synthetic Negatives PU + PN compatible Object-centric FPR eval

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 and integrate them into vision-based traversability learning as a training strategy that plugs into both Positive–Unlabeled (PU) and Positive–Negative (PN) frameworks without modifying inference architectures. We also introduce an object-centric FPR evaluation that measures errors specifically on the inserted negative regions, providing an annotation-free indicator of non-traversable recognition.

SyNeT abstract overview

Key idea of SyNeT: explicit synthetic negatives for traversability learning.

Method

SyNeT method pipeline (replace this file)

Synthetic Negative Generation Module

SyNeT method pipeline (replace this file)

SyNeT integrated into PU and PN baselines: synthetic negatives provide explicit non-traversable features that enforce clearer separation in the embedding space.

1) Synthetic Negative Generation

  • Region selection: sample ROI and target size inside the ground area.
  • Inpainting: generate scene-consistent negatives with diffusion inpainting (e.g., Stable Diffusion 3.5 / FLUX.1 Fill).
  • Segmentation & filtering: segment the generated object and apply scene filters (object count, pixel area); resample if violated.
  • Composition: blend the approved negative into the base image to obtain a composite and a pixel-wise negative mask.

2) Training Strategy (Plug-in)

  • PU (LORT): add a negative-center assignment loss for synthetic negatives + unlabeled features, plus a repulsion loss to avoid center collapse.
  • PN (V-STRONG): add a contrastive loss term that treats synthetic negatives as reliable negative anchors.
  • No inference change: SyNeT modifies only training objectives; at test time, you run the original model and produce a similarity map.

Qualitative Results

Below are example traversability maps across diverse environments (social, indoor, off-road, and urban).

Qualitative result 1 (replace this file)

Video

Citation

If you find this useful, please cite:

        ------

Update the BibTeX fields once the author list / venue / year are final.