Introduction
Consumer-grade drones, such as the DJI Mini 3 Pro, offer high portability but lack the specialized hardware—like LiDAR or dedicated depth sensors—typically found in research-grade platforms. Most autonomous exploration algorithms assume these sensors are present along with significant onboard compute. This paper addresses the challenge of performing high-quality mapping and efficient multi-UAV exploration using only monocular RGB cameras and limited processing power.
System Architecture and Methods
The proposed pipeline adapts the FUEL and RACER frameworks to handle the constraints of consumer hardware. A key innovation is the shift from single viewpoint planning to viewpoint pair sampling. Since monocular drones require motion to estimate depth via triangulation, the system plans for pairs of related positions. These pairs are evaluated using a heuristic that considers distance, parallax, and overlap to ensure stable depth estimation using the MASt3R transformer-based model.
To maintain visual-inertial odometry (VIO) and avoid "blind" movements, a yaw-aware A* planner is introduced. This planner ensures the drone's heading remains aligned with its velocity vector within the camera's field of view. Furthermore, to overcome computational bottlenecks, the authors employ a semi-distributed communication scheme. Global mapping and coordination are offloaded to a central base node (e.g., a ground station), while individual UAVs handle local trajectory execution and refinement.
Experimental Results
The system was evaluated in the Flightmare simulator, a high-fidelity Unity-based engine for quadrotor research. The results demonstrate that the heuristic-based viewpoint pairing significantly improves map quality compared to naive sampling strategies.
As shown in the table above, the heuristic method achieved an Average Hausdorff Distance (AHD) of 0.30m and a reconstruction accuracy (ACC) of 93%, vastly outperforming the baseline "no-pairing" approach (0.68m AHD and 53% ACC).
In terms of trajectory reliability, the yaw-aware A* planner reduced the Absolute Trajectory Error (ATE) to 0.14m, compared to 0.26m for a standard constrained planner. Multi-UAV experiments showed that the semi-distributed coordination reduced total path length by approximately 25% and energy consumption by 24% compared to fully decentralized alternatives like RACER.
Conclusion
By optimizing for the specific sensing requirements of monocular cameras and utilizing a semi-distributed compute model, this work brings advanced autonomous exploration to consumer-grade hardware. While currently validated in simulation, future work will focus on real-world deployment on DJI platforms to test robustness against environmental noise and communication latency.
Summary updated