TL;DR
The LUOJIA Explorer is an autonomous drone designed to map and navigate "unexposed spaces"—hazardous, GNSS-denied environments like caves, tunnels, and indoor facilities. By combining solid-state LiDAR with a computationally efficient, ESDF-free planning algorithm, the system achieves stable self-positioning and high-speed exploration without relying on external signals.
What Problem Does This Paper Address?
Exploring "unexposed spaces" (underground areas, karst caves, or complex indoor environments) is a high-stakes task for humans due to extreme lighting variations, irregular geometries, and the total absence of GPS/GNSS signals. While Unmanned Aerial Vehicles (UAVs) are ideal for these environments due to their agility, they face significant technical hurdles:
- Signal Denial: Traditional navigation fails without GPS.
- Computational Constraints: Standard autonomous planning often relies on Euclidean Symbolic Distance Fields (ESDF) to avoid obstacles, which requires heavy processing power that can lag during high-speed flight.
- Sensor Limitations: Many visual sensors struggle in low-light or featureless environments.
The authors introduce the LUOJIA Explorer to bridge these gaps, providing a robust hardware-software integration for "blind" autonomous exploration.
Methods
The LUOJIA Explorer utilizes a specialized hardware stack and a two-part software subsystem:
1. Hardware Integration
The drone is built on a UAV450 frame equipped with a DJI Livox MID-360 solid-state LiDAR (which provides a wide field of view and built-in IMU) and an external anti-collision cage to protect the rotors during tight-quarter maneuvers.
2. Localization and Mapping Subsystem
The system uses an Iterative Kalman Filter (IKF) to fuse LiDAR point clouds with IMU data.
- Efficiency Hack: Instead of calculating Kalman gain based on the number of measurements (which is huge in point clouds), they calculate it based on the state dimension. This drastically reduces the math required per frame.
- Map Maintenance: It uses an incremental k-d tree structure for real-time map updates and a "distance-based thinning" module to keep the navigation map lightweight.
3. ESDF-Free Planning Subsystem
Most drones pre-calculate a 3D "distance map" (ESDF) of their surroundings to find safe paths. LUOJIA Explorer skips this. Instead, it uses:
- A* Search: To find an initial "bootstrap" path.
- Gradient-Based Optimization: It refines the path using B-splines, applying "collision penalties" only when a trajectory actually hits an obstacle. This "lazy" approach saves massive amounts of CPU time, allowing for faster flight.
Datasets
The researchers evaluated the system in two primary settings:
- Simulation: Conducted using the XTDrone platform in a software-in-the-loop (SITL) configuration, simulating an outdoor environment where GNSS signals were intentionally disabled.
- Real-World Testbed: A controlled "unexposed space" at Wuhan University measuring 6.28m x 4.74m x 3m, filled with obstacles like nets and furniture to test obstacle avoidance and mapping accuracy.
Key Insights and Ideas
- The "Lazy" Planner is Faster: By moving away from ESDF-based planning, the authors demonstrate that you don't need to know the distance to every obstacle at all times—you only need to know when your current path is in danger. This allows the drone to react more quickly to new sensor data.
- State-Space Optimization: Shifting the Kalman Filter math from the measurement dimension to the state dimension is a critical optimization for LiDAR-heavy systems, preventing the onboard computer from becoming a bottleneck.
- Robustness through Redundancy: The integration of a physical anti-collision structure acknowledges the reality of exploring unknown spaces: even the best algorithms might have a close brush with a wall in a tight cave.
Key Results
- Exploration Efficiency: The system achieved an exploration efficiency of 88.83 m³/s.
- Speed: In real-world tests, the UAV autonomously navigated to a target point while avoiding obstacles in just 15 seconds.
- Mapping Latency: The system completed the construction of the experimental site map in approximately 1 second.
- Stability: Both simulation and real-world flights demonstrated stable self-positioning and zero-drift localization in environments where GPS was completely unavailable.
Limitations and Future Work
While the results are promising, the authors note a few areas for growth:
- Scale: The real-world experiments were conducted in a relatively small indoor lab. Testing in larger, more complex natural environments (like actual karst caves) is a necessary next step.
- Target-Based Exploration: Currently, the drone explores generally; the authors plan to implement "target-based" logic where the UAV searches for specific objects or features.
- Multi-Agent Collaboration: There is potential to expand this into an air-ground cooperative system, where the UAV maps from above while ground robots handle heavy-duty tasks below.
Summary updated