GSAT: Geometric Traversability Estimation Using Self-Supervised Learning with Anomaly Detection for Diverse Terrains
Dongjin Cho, Miryeong Park, Juhui Lee, Geonmo Yang, Younggun Cho
AI summary
Problem
Self-supervised traversability estimation struggles with the positive-only learning problem, where the absence of negative samples causes unstable feature representations and unreliable decision boundaries. Traditional methods also depend on subjective human-defined thresholds that fail to generalize across diverse, unstructured terrains.
Approach
The method constructs a positive hypersphere in latent space to separate normal from anomalous regions using only positive traversal data. It jointly optimizes anomaly detection and traversability prediction while applying targeted geometric augmentations to increase terrain diversity.
Key results
- Constructs a stable positive hypersphere boundary without auxiliary prototypes
- Jointly learns anomaly detection and traversability prediction for efficient terrain assessment
- Introduces geometric augmentations (flipping, yaw, pitch) to mitigate directional and slope biases
- Achieves superior anomaly classification and traversability estimation on RELLIS-3D and DITER++ datasets
Why it matters
Enables robots to safely navigate unstructured environments autonomously by learning robust traversability boundaries directly from experience, reducing reliance on manual tuning and labeled data.
Abstract
Safe autonomous navigation requires reliable esti- mation of environmental traversability. Traditional methods have 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 ex- plicit 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 requiring additional prototypes (e.g., unlabeled or negative). Furthermore, our approach employs joint learning of anomaly classification and traversability prediction to more efficiently utilize robot experience. We comprehensively evaluate the proposed framework through ablation studies, validation on heterogeneous real-world robotic platforms, and autonomous nav- igation demonstrations in simulation environments. Our method is available at https://sparolab.github.io/research/ gsat/.