MonoSpheres: Large-Scale Monocular SLAM-Based UAV Exploration through Perception-Coupled Mapping and Planning

Technical Summary of MonoSpheres: Large-Scale Monocular SLAM-Based UAV Exploration

Problem Statement
Autonomous exploration of unknown 3D environments is a critical capability for Uncrewed Aerial Vehicles (UAVs) in applications such as search-and-rescue and planetary exploration. While significant progress has been made with robots equipped with dense range sensors (e.g., LiDAR, RGB-D cameras), these sensors are often heavy, expensive, and limit flight time. Conversely, monocular cameras and IMUs are lightweight and cost-effective but present a fundamental challenge: motion-based depth estimation (as used in Visual SLAM) produces sparse, texture-dependent depth data with high uncertainty in textureless areas. Existing monocular exploration methods have largely been restricted to small-scale indoor environments or single rooms because they rely on assumptions valid for dense sensors (e.g., that non-detection implies free space) which fail with sparse data, leading to map gaps, incorrect free-space initialization, and collisions.

Methodology
The authors propose MonoSpheres, a perception-coupled mapping and planning framework designed explicitly for the properties of sparse monocular depth sensing. The system operates on a graph-of-spheres representation rather than traditional voxel grids or TSDFs.

1. Mapping Module:

   • Representation: The map consists of a graph of intersecting spheres representing free space and a set of 3D points representing obstacles.
   • Depth Interpolation & Polyhedron Construction: The system constructs a "depth mesh" from tracked SLAM points using Delaunay triangulation to define a visible free-space polyhedron ($P_f$) between the camera and the observed points.
   • Open-Area Virtual-Depth Estimation (OVDE): To address the inability of SLAM to estimate depth in textureless open areas (e.g., sky or flat ground), the method injects "virtual points" into the depth mesh. If a virtual point remains unobstructed across multiple poses with sufficient parallax, it is treated as free space, allowing the UAV to fly over open areas without gaps in the map.
   • Distance-Based Obstacle Filtering (DBOF): To prevent the deletion of mapped obstacles when they are momentarily occluded or not tracked, the system employs a distance-based filtering logic. A map point is only deleted if it falls within the visible free-space polyhedron and is closer to the camera than a threshold relative to its measurement distance. This preserves "protected points" that constrain sphere radii, preventing the map from erroneously erasing obstacles.

2. Planning and Exploration Module:

   • Frontier Sampling: Frontiers are sampled along the boundary of the visible free-space polyhedron rather than an occupancy grid.
   • Perception-Aware Heading Control: Recognizing that monocular depth requires translational motion to be estimated, the system enforces a Forward-Facing Flight (FFF) strategy. When approaching an exploration viewpoint, the UAV aligns its heading with the goal direction to ensure pure translational motion, maximizing the likelihood of observing parallax and estimating depth.
   • Viewpoint Blocking: To prevent the UAV from wasting time on textureless surfaces where no new depth can be estimated, the system blocks the sampling of new viewpoints near previously visited ones if no new frontiers were discovered.

Key Contributions

• Novel Mapping Approach: A method that explicitly interpolates depth and estimates free space in texture-sparse areas while tracking obstacle uncertainty, enabling the construction of large-scale maps from sparse monocular data.
• Perception-Coupled Planning: An exploration strategy that integrates the limitations of monocular sensing (need for translation, texture dependence) directly into the control loop via heading alignment and viewpoint blocking.
• First Real-World Outdoor Demonstration: To the authors' knowledge, this is the first method to achieve 3D monocular exploration in real-world unstructured outdoor environments, moving beyond the scale of single indoor rooms.
• Open-Source Benchmark: The release of code and simulation scripts to facilitate future research in monocular exploration.

Experimental Results
The system was evaluated in diverse real-world scenarios (an abandoned farm and an orchard) and simulated environments (Earthquake, Cave, Rooftops, and Sparse worlds).

• Real-World Performance: The UAV successfully autonomously explored a 70x10x8m area around an abandoned farmhouse and a cluttered orchard, mapping obstacles, avoiding debris, and returning to base upon low battery.
• Simulation Performance: In comparative studies against a grid-based ray-tracing baseline (similar to prior works [4], [5]), MonoSpheres consistently explored significantly larger areas and volumes across all test worlds.
• Ablation Studies: Removing key components led to failure:
  • Without OVDE, the UAV could not explore open areas and remained trapped in the starting "valley."
  • Without DBOF, the UAV overwrote mapped obstacles with free space, leading to collisions.
  • Without FFF, the UAV failed to detect obstacles due to insufficient parallax, resulting in crashes.
  • Without Viewpoint Blocking, performance degraded due to redundant exploration of textureless surfaces.

Significance and Limitations
The paper claims that by redesigning both mapping and planning to account for the specific regularities of motion-based depth sensing (distance-dependent accuracy, texture dependence, and trajectory dependence), MonoSpheres achieves autonomy at scales comparable to dense sensing methods without the associated hardware costs. The authors note that the system's primary limitation remains the detection capabilities of the underlying visual SLAM frontend; it cannot map obstacles that the frontend fails to detect (e.g., thin wires or large textureless objects). Future work suggests integrating low-cost ultrasonic sensors or switching to dense SLAM frontends to address these specific detection gaps.