MARSIM: A light-weight point-realistic simulator for LiDAR-based UAVs

Fanze Kong, Xiyuan Liu, Benxu Tang, Jiarong Lin, Yunfan Ren, Yixi Cai, Fangcheng Zhu, Nan Chen, Fu Zhang

Introduction

Simulating unmanned aerial vehicles (UAVs) is critical for developing navigation algorithms without the risk of physical hardware damage. However, traditional simulators like Gazebo or AirSim often struggle to replicate the complexity of real-world environments. These tools typically rely on 3D mesh models, which are difficult to generate accurately from real-world scans, often leading to artifacts like "closed" surfaces where gaps should exist or missing details due to occlusions.

Figure 1: MARSIM enables point-realistic simulation by rendering LiDAR scans directly from raw 3D point cloud data, bridging the gap between virtual environments and real-world complexity. (a) Global point cloud map, (b) simulated scan, (c) real-world reference.

MARSIM addresses these challenges by introducing a "point-realistic" simulation approach. Developed by researchers at the University of Hong Kong, MARSIM bypasses the mesh-reconstruction phase entirely, rendering realistic LiDAR scans directly from raw 3D point cloud data. This approach preserves fine environmental details and significantly reduces the computational overhead associated with dense mesh rendering.

System Architecture and Methods

The core of MARSIM is a high-performance, ROS-compatible rendering pipeline designed for efficiency and realism. Unlike traditional ray-casting methods, MARSIM utilizes a point-to-scan rendering process.

Figure 2: Overview of the MARSIM system architecture, highlighting the integration with ROS and the modular components for dynamics, LiDAR simulation, and planning.

The rendering pipeline consists of several key stages:

  1. Preprocessing: Global point cloud maps are imported, spatially downsampled, and organized into voxels for rapid lookup.
  2. Depth Image Construction: Map points within the UAV’s Field of View (FoV) are projected onto a dense depth image.
  3. Occlusion Culling & Interpolation: To manage point cloud sparsity, MARSIM employs a "smallest depth wins" strategy and calculates interpolation ranges to ensure foreground objects correctly occlude the background.
  4. Plane Correction: Map points are treated as small local planes to prevent artifacts on slanted surfaces, ensuring accurate laser ray intersections.
  5. Pattern Masking: The final depth image is masked with specific LiDAR scanning patterns, such as the non-repetitive flower-like pattern of the Livox Avia.

Benchmark Datasets

To evaluate the simulator, the authors developed a benchmark suite comprising 10 high-resolution (0.01m) real-world point cloud maps. These environments include natural forests, complex indoor workshops, multi-layer corridors, and infrastructure like parking garages and the historic HKU main building. These maps provide challenging scenarios featuring cluttered obstacles, narrow passages, and thin structures that are typically difficult to represent with meshes.

Experimental Results

MARSIM demonstrates significant improvements in performance and resource efficiency compared to industry standards:

  • Computational Efficiency: The GPU-accelerated version of MARSIM is up to 10x faster than Gazebo in high-resolution scenes. Even the CPU-only version outperforms Gazebo's GPU-accelerated rendering.
  • Memory Usage: MARSIM requires approximately half the RAM of Gazebo for equivalent environments.
  • Real-time Performance: On consumer-grade hardware (e.g., an Intel NUC i7), the simulator maintains a rendering rate exceeding 10 Hz, even with maps containing tens of millions of points.
  • Sim-to-Real Fidelity: By utilizing the same point cloud maps that a drone would generate in the real world, simulated flight trajectories closely match experimental data.

Conclusion and Future Work

MARSIM provides a robust, light-weight solution for high-fidelity UAV simulation. By directly utilizing point cloud data, it preserves environmental details and enables real-time testing on mobile computing platforms. Future work will focus on optimizing the rendering pipeline for very sparse LiDAR patterns and expanding support for complex dynamic agents, such as non-rigid bodies and walking humans.

Summary updated

BibTeX
@article{kong2023marsim,
  author = {Fanze Kong and Xiyuan Liu and Benxu Tang and Jiarong Lin and Yunfan Ren and Yixi Cai and Fangcheng Zhu and Nan Chen and Fu Zhang},
  title = {{MARSIM: A light-weight point-realistic simulator for LiDAR-based UAVs}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2023},
  url = {https://openreview.net/forum?id=C4lzzhtbmH},
}
Figure