LiDARFlow: Real-Time Panel-Based MAV Guidance in Unknown Environments
This preprint describes a lightweight onboard guidance system for micro aerial vehicles in unknown environments, combining a guiding vector field for nominal motion with panel-method potential-flow obstacle avoidance. In indoor flight tests at 0.5 m/s, a quadrotor using only a Livox MID-360 LiDAR and Raspberry Pi 5 successfully completed waypoint-navigation and direction-following tasks while avoiding all obstacles. The result matters because it shows map-free sense-and-avoid can run in real time on constrained onboard hardware, reducing operator workload for cluttered or emergency missions.
Can a computationally lightweight onboard system that fuses a nominal guiding vector field with panel-based potential-flow obstacle avoidance, using only online LiDAR perception, enable a micro aerial vehicle to navigate unknown cluttered environments safely and in real time?
MAVs often operate in unknown, cluttered environments where no prior map exists, and pilots may lack extensive experience. Existing optimization-based planners typically require building and maintaining maps and can be too computationally demanding for small onboard computers, while classical reactive methods such as potential fields are prone to local minima, oscillation, and lack of continuous-time safety guarantees.
Prior approaches include optimization-based planners such as FASTER, MADER, EGO-Planner, and EGO-Swarm, which can be expensive and map-dependent, and reactive methods including control barrier functions, velocity obstacles/ORCA, and artificial potential fields, which trade guarantees for speed. The PGFlow panel method itself previously assumed complete prior knowledge of obstacle geometry, and earlier unknown-environment extensions were demonstrated only on a ground robot or in simulation.
The proposed system superposes a convergent guiding vector field for nominal motion with the harmonic potential-flow PGFlow obstacle-avoidance method. Obstacles are not known a priori; they are extracted online from Livox MID-360 LiDAR point clouds through voxel downsampling, region-of-interest filtering, clustering, iterative RANSAC line fitting, and a temporal segment map. The extracted convex obstacle panels are inflated by half the drone size and fed to PGFlow to generate repulsive velocity fields in the horizontal plane. The full pipeline runs onboard a Raspberry Pi 5 and sends velocity commands to a Paparazzi autopilot.
In two indoor flight scenarios at a maximum speed of 0.5 m/s, the vehicle successfully reached both waypoints and returned in the waypoint task, and successfully tracked commanded directions of (0,1) then (1,0) in the directional task, avoiding all placed obstacles in each case. Obstacle positions were unknown to the drone. PGFlow guidance ran at 100 Hz; the maximum allowed panel count decreased from 2,025 panels at 1 Hz to 1,100 at 5 Hz, 725 at 10 Hz, and 506 at 20 Hz. The paper does not report quantitative success rates or baseline comparisons.
The authors state that trajectories showed noticeable oscillations and sharp turns, attributed to unstable obstacle detection where estimated obstacle dimensions changed erratically between frames. Testing was limited to low speed, static convex obstacles, a single MAV, and an external OptiTrack motion capture system rather than onboard state estimation. Guidance was handled only in the horizontal plane. Concave obstacles were not tested and may introduce local minima. The point cloud was transmitted from a C++ driver to a Python guidance module before processing, adding communication overhead; the O(n^3) panel matrix inversion limits scaling with the number of panels. No code release or multi-flight quantitative evaluation is reported.
This could inform lightweight onboard sense-and-avoid modules for small inspection drones, search-and-rescue UAVs, and indoor industrial MAVs, especially where operators are not expert pilots. Useful deployment is likely 1–3 years away and would first require onboard state estimation, higher-speed validation, dynamic and concave obstacle handling, more stable point-cloud segmentation, and multi-vehicle testing.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.