Real-time Spatial-temporal Traversability Assessment via Feature-based Sparse Gaussian Process

Senming Tan, Zhenyu Hou, Zhihao Zhang, Long Xu, Mengke Zhang, Zhaoqi He, Chao Xu, Fei Gao, Yanjun Cao

TL;DR

Researchers from Zhejiang University have developed FSGP-BGK, a high-efficiency framework for robot navigation that assesses terrain traversability in real-time. By combining feature-based Sparse Gaussian Processes (SGP) with a Bayesian Gaussian Kernel (BGK) fusion method, the system creates smooth, accurate maps that integrate both immediate sensor data and historical observations, outperforming state-of-the-art methods in both accuracy and computational speed.

What Problem Does This Paper Address?

For a mobile robot to navigate an outdoor, unstructured environment—like a forest or a ruin—it must determine which areas are "traversable" (safe to drive on) and which are hazardous.

Traditional methods face a "trilemma" of challenges:

  1. Computational Cost: Standard Gaussian Processes (GP) provide excellent uncertainty estimates but are mathematically "heavy" ($O(N^3)$ complexity), making them too slow for real-time use as data accumulates.
  2. Information Loss: Simple elevation maps often ignore critical geometric features like local curvature or steep gradients that define obstacles.
  3. Temporal Blindness: Many existing Sparse GP (SGP) methods only look at the current "frame" of sensor data. They fail to effectively remember what the robot saw a few seconds ago, leading to flickering or incomplete maps when sensors are occluded.

Methods

The authors propose a pipeline called FSGP-BGK to solve these issues:

  • Feature-Based Extraction: Instead of just looking at height, the system extracts curvature and gradient features directly from raw point clouds. It uses GPU-accelerated K-Nearest Neighbors (KNN) and Principal Component Analysis (PCA) to identify "feature points" (like the edges of a cliff) while downsampling flat, redundant areas.
  • Sparse Gaussian Process (SGP): To keep the math fast, the system uses "inducing points"—a small subset of representative data points—to represent the entire terrain. This allows for continuous, smooth regression of the ground surface without the $O(N^3)$ slowdown.
  • Spatial-Temporal BGK Fusion: This is the "secret sauce." Instead of just piling up old point clouds (which would eventually crash the computer), they use a Bayesian Gaussian Kernel to fuse the current SGP prediction with a historical "memory" buffer. This ensures the map remains consistent over time and fills in gaps caused by sensor shadows.
  • GPU Acceleration: The entire pipeline—from feature extraction to the final traversability score—is optimized for the GPU, allowing the robot to update its map at high frequencies (20Hz).

Datasets

The researchers evaluated their system using a mix of simulated and real-world data:

  • EPFL Terrain Generator: Used to synthesize large-scale (50m x 50m) Hilly, Forest, and Ruin environments.
  • Complex Urban Dataset: Used to simulate realistic on-road urban settings.
  • SceneNet: Used for indoor room environments.
  • Real-World Testing: Validated on an Agilex Scout Mini differential-drive robot equipped with a Livox MID-360 LiDAR and an onboard NUC11 i7 computer.

Key Insights and Ideas

  • Features over Raw Data: The paper argues that curvature and gradient are better "proxies" for navigability than raw elevation. By prioritizing these features during the SGP training phase, the model captures critical obstacles with fewer data points.
  • Efficient Memory: The BGK fusion acts as a smart filter. It uses a temporal decay weight (forgetting very old data) and an uncertainty-based confidence weight. This allows the robot to maintain a "global-like" understanding of its surroundings using only local processing power.
  • Decoupling CPU/GPU: By moving the heavy lifting (KNN, PCA, SGP) to the GPU, the CPU is freed up for high-level path planning and control, which is vital for stable autonomous flight or driving.

Key Results

  • Accuracy: In outdoor forest scenarios, FSGP-BGK improved mean accuracy by 18% to 20% over baseline SGP. In indoor environments, the improvement was even more dramatic, reaching 46% to 48%.
  • Efficiency: The system achieved an average runtime of 33.84ms, roughly 3x faster than traditional Elevation Map (EM) methods (107.85ms).
  • Memory Savings: Compared to simply accumulating multi-frame data, the BGK method reduced GPU memory consumption from 36% to 25% and cut processing time by roughly 30%.
  • Robustness: Real-world tests showed the system could identify small obstacles that baseline SGP methods missed, enabling the robot to navigate complex, undulating terrain without collisions.

Limitations and Future Work

While the system excels at local traversability, the authors note a few areas for growth:

  • Global Mapping: The current framework is "global-map-free," focusing on the robot's immediate vicinity. Integrating this into a persistent global SLAM (Simultaneous Localization and Mapping) framework could improve long-range mission planning.
  • Multi-Robot Systems: The authors suggest extending this approach to multi-robot exploration, where robots could share their spatial-temporal traversability "memories" to map large areas faster.
  • Dynamic Obstacles: While the temporal decay helps, explicitly modeling moving objects (like humans or other vehicles) within the SGP framework remains a challenge for future iterations.

Check out the code: The authors have open-sourced their implementation at https://github.com/ZJU-FAST-Lab/FSGP_BGK.

Summary updated

Abstract

Terrain analysis is critical for the practical application of ground mobile robots in real-world tasks, especially in outdoor unstructured environments. In this paper, we propose a novel spatial-temporal traversability assessment method, which aims to enable autonomous robots to effectively navigate through complex terrains. Our approach utilizes sparse Gaussian processes (SGP) to extract geometric features (curvature, gradient, elevation, etc.) directly from point cloud scans. These features are then used to construct a high-resolution local traversability map. Then, we design a spatial-temporal Bayesian Gaussian kernel (BGK) inference method to dynamically evaluate traversability scores, integrating historical and real-time data while considering factors such as slope, flatness, gradient, and uncertainty metrics. GPU acceleration is applied in the feature extraction step, and the system achieves real-time performance. Extensive simulation experiments across diverse terrain scenarios demonstrate that our method outperforms SOTA approaches in both accuracy and computational efficiency. Additionally, we develop an autonomous navigation framework integrated with the traversability map and validate it with a differential driven vehicle in complex outdoor environments. Our code will be open-source for further research and development by the community, https://github.com/ZJU-FAST-Lab/FSGP_BGK.
BibTeX
@inproceedings{tan2025real,
  author = {Senming Tan and Zhenyu Hou and Zhihao Zhang and Long Xu and Mengke Zhang and Zhaoqi He and Chao Xu and Fei Gao and Yanjun Cao},
  title = {{Real-time Spatial-temporal Traversability Assessment via Feature-based Sparse Gaussian Process}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11246696},
}

Loading related content...

Figure