IEEE ICRA 2026

GSAT: Geometric Traversability Estimation using Self-supervised Learning with Anomaly Detection for Diverse Terrains

Robot-specific traversability from LiDAR — without manual labels — by learning a positive hypersphere in latent space.

Dongjin Cho1, Miryeong Park1, Juhui Lee1, Geonmo Yang1, and Younggun Cho1†

1Spatial AI and Robotics (SPARO) Lab, Inha University, South Korea

Paper arXiv Video Code
Background

Motivation

Reliable traversability is essential for safe navigation, yet existing pipelines still rely heavily on human-defined rules or struggle with the positive-only nature of robot experience.

Human-Defined Approaches

Geometric feature-based traversability estimation
  • Point clouds are projected into elevation maps, from which geometric attributes are extracted.
  • Traversability scores are assigned via hand-crafted thresholds based on the robot's physical constraints.
Semantic class mapping for traversability
  • Semantic. Classes are mapped to predefined Safe / Dangerous categories.
  • Traversability tables are manually defined, requiring scene-specific tuning.

Limitation

  • Inaccurate predictions due to subjective human supervision
  • Heavy manual effort for threshold tuning and label definition

Self-supervised Learning & Remaining Challenges

To avoid human supervision, self-supervised methods learn traversability from the robot's own traversal experience. However, their main challenge is the positive-only learning problem: without negative examples, contrastive samples must be constructed to distinguish traversable from non-traversable regions.

Existing Works: Anomaly Detection for Positive-Only Learning

Reconstruction-based self-supervised learning
  • Reconstruction-based. Trains the model to reconstruct only the positive samples.
  • Identifies anomalies using high reconstruction error.
Prototype-based positive-only learning
  • Prototype-based. Learns prototypes from unlabeled or negative samples for contrastive comparison.
  • Using unlabeled prototypes can introduce feature distortion, which degrades fine-tuning performance.
  • Generating negative samples typically requires large pre-trained models.

Remaining Challenges

  • Strong dependence on heuristic design
  • Feature distortion caused by unlabeled data
  • Lack of a geometric foundation model for traversability
GSAT traversability estimation overview

TL;DR — GSAT does not require any negative samples or unlabeled prototypes.

Overview

Abstract

Safe autonomous navigation requires reliable estimation of environmental traversability. Prior research has relied on semantic or geometry-based approaches with human-defined thresholds, but these methods often yield unreliable predictions due to the inherent subjectivity of human supervision. While self-supervised approaches enable robots to learn from their own experience, they still face a fundamental challenge: the positive-only learning problem.

To address these limitations, recent studies have employed Positive-Unlabeled (PU) learning, where the core challenge is identifying positive samples without explicit negative supervision. In this work, we propose GSAT, which addresses these limitations by constructing a positive hypersphere in latent space to classify traversable regions through anomaly detection without explicit negative samples.

Furthermore, our approach employs joint learning of anomaly classification and traversability prediction to more effectively utilize robot experience. We comprehensively evaluate the proposed framework through ablation studies, validation on heterogeneous real-world robotic platforms (Wheeled and Legged), and autonomous navigation demonstrations in complex simulation environments.

Positive Hypersphere

Latent anomaly detection clusters familiar terrain without explicit negative supervision.

Joint Learning

Anomaly, reconstruction, and regression losses refine traversability from robot experience.

Platform-Aware Maps

Wheeled and legged robots receive distinct traversability maps aligned with mobility constraints.

Approach

Method Overview

A

Automated Data Generation

SLAM trajectories projected to BEV grids for efficient real-time supervision.

B

Pillar-Based Encoding

Deep spatial features encoded into latent terrain representations.

C

Experience-Aware Learning

Anomaly detection separates normal/anomalous samples to refine the hypersphere.

GSAT overall pipeline
Figure: Overall architecture — automated data generation, pillar-based encoding, and joint learning.

The proposed GSAT framework systematically learns robot-specific traversability through three main components:

  • (A) Automated Data Generation: Generates supervision by projecting SLAM trajectories into spatial BEV grids from 3D LiDAR point clouds.
  • (B) Traversability Network: Pillar-based architecture extracts deep spatial features into a latent terrain representation.
  • (C) Experience-Aware Traversability Learning: Anomaly detection in latent space separates unlabeled data; joint optimization via anomaly, reconstruction, and regression losses.

By integrating these components, GSAT clusters familiar terrains while isolating unseen or hazardous regions — enabling safe navigation without manual labeling.

Results

Experimental Evaluation

Quantitative Analysis: Anomaly Loss Formulation

We analyze how different handling of unlabeled data within the anomaly loss affects traversability estimation across RELLIS-3D and DITER++.

  • w/o Anomaly Loss: Deep-SVDD suffers from inflated recall but fails to distinguish anomalous features.
  • All Unlabeled as Anomalous: Overly restrictive hypersphere with poor generalization.
  • Anomalous-only: Better precision-recall balance but lacks guidance to cluster normal samples.
Proposed (Normal/Anomalous): Highest F1-scores (77.61%, 88.04%) — incorporating normal samples is crucial for robust boundary refinement.
Ablation results table
Table 1: Anomaly loss configurations on RELLIS-3D and DITER++.
Citation

BibTeX

@inproceedings{cho2026gsat,
  author    = {Cho, Dongjin and Park, Miryeong and Lee, Juhui and Yang, Geonmo and Cho, Younggun},
  title     = {GSAT: Geometric Traversability Estimation using Self-supervised Learning with Anomaly Detection for Diverse Terrains},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026},
}