Abstract
This paper presents a system that enables delivery drones to autonomously navigate and deliver packages to specific locations around a house, such as a front door or backyard, without the need for external markers or specialized landing pads. By leveraging deep learning for semantic scene understanding and vision-based navigation, the system identifies safe descent zones and routes the drone through the "last-meter" of the delivery process. In simulated evaluations across 20 house layouts, the proposed system reached its destination 161% faster than traditional frontier exploration-based strategies.
Introduction
While drone delivery technology has advanced significantly, the final stage of delivery—getting the package from a general GPS coordinate to the recipient's actual doorstep—remains a challenge. Most current commercial solutions rely on pre-installed infrastructure like visual markers or landing pads. This work addresses the "last-meter" problem by enabling drones to interpret the unstructured environment of a residential home using only onboard sensors, much like a human courier would.
System Architecture
The authors propose a two-stage autonomous pipeline designed to handle the transition from high-altitude flight to precise doorstep delivery. This pipeline is divided into a descent module and a navigation module.
1. Descending to the House
Upon reaching the target GPS coordinates, the drone utilizes a downward-facing camera and a Unet-based semantic segmentation model. This model parses the scene into categories such as roofs, grass, and paved areas.
- Orientation Estimation: A key innovation is the use of semantic cues to infer house orientation. By identifying the shared boundary between the roof and the driveway or pavement, the system determines the likely "front" of the house.
- Safe Descent Spot: To mitigate GPS drift and ensure safety, the system calculates a descent coordinate that maintains a 2.5-meter buffer from the house footprint and other obstacles like trees or vehicles.
2. Vision-Based Navigation to the Door
Once the drone descends to a height of approximately 2 meters, it switches to a front-facing stereo camera for the final approach.
- Mapping and Detection: The system employs RTAB-Map to generate a 3D OctoMap for real-time obstacle avoidance. Simultaneously, a YOLOv3 model is used to detect the front door.
- Footprint-Guided Search: Rather than exploring the environment randomly, the drone follows a path along the house's "footprint" derived from the earlier aerial segmentation. This guided search significantly reduces the time required to locate the delivery site.
Evaluation and Results
The system was trained using the Semantic Drone Dataset for aerial parsing and the DoorDetect Dataset for door identification. Testing was conducted in AirSim, a high-fidelity simulator, across 20 distinct residential environments.
Key performance metrics include:
- Efficiency: The drone reached the front door in an average of 89.45 seconds, compared to 144 seconds for the frontier exploration baseline.
- Speedup: The proposed method demonstrated a 161% improvement in speed over the baseline.
- Reliability: The system successfully differentiated between front and back yards based on user preference in all test cases, demonstrating robust spatial reasoning.
Conclusion and Future Work
This research demonstrates that semantic scene understanding can replace the need for physical markers in drone delivery. By using environmental context to infer orientation and guide search patterns, drones can navigate complex residential areas more efficiently than through general exploration.
Future work will focus on improving the system's robustness to environmental factors like shadows and varying lighting conditions, which can occasionally interfere with semantic segmentation. Additionally, the authors aim to integrate natural language processing to allow users to provide specific delivery instructions, such as "leave the package by the side gate."
Summary updated