Abstract In this work, we present a system that enables delivery drones to autonomously navigate and deliver packages at various locations around a house according to the desire of the recipient and without the need for any external markers as currently used. This development is motivated by recent advancements in deep learning that can potentially supplant the specialized markers presently used by delivery drones for identifying sites at which to deliver packages. The proposed system is more natural in that it takes instruction on where to deliver the package as input, similar to the instructions provided to human couriers. First, we propose a semantic image segmentation-based descending location estimator that enables the drone to find a safe spot around the house at which it can descend from higher altitudes. Following this, we propose a strategy for visually routing the drone from the descent location to a specific site at which it is to deliver the package, such as the front door. We extensively evaluate this approach in a simulated environment and demonstrate that with our system, a delivery drone can deliver a package to the front door and also to other specified locations around a house. Relative to a frontier exploration-based strategy, drones using the proposed system found and reached the front doors of the 20 test houses 161% faster.
Introduction
While drone delivery is no longer science fiction, the "last-meter" remains a significant hurdle. Current commercial solutions often rely on specialized landing pads, external markers like QR codes, or lowering packages via tether to a general GPS coordinate. However, recipients typically prefer packages delivered to specific locations, such as a front door or side yard—spots that GPS alone cannot hit with sufficient precision. Furthermore, requiring customers to place physical markers is a barrier to mass adoption. This research addresses the challenge of marker-free, semantic-aware delivery, enabling a drone to understand property layouts and navigate safely to user-requested locations.
Methodology
The proposed system utilizes a two-stage pipeline to transition the drone from a high-altitude GPS coordinate to a specific delivery site.
System Architecture
System architecture for autonomous drone delivery. Source, p. 3
- Semantic Descending Location Estimation: Using a downward-facing camera, the system performs semantic segmentation to identify roofs, paved areas, and grass. It calculates house orientation by analyzing the boundary between the roof and paved regions, allowing it to classify yard areas. The system then searches for a descent spot at least 2.5 meters away from obstacles and the house footprint.
- Vision-Based Navigation to Door: Upon descending to approximately 2 meters, the drone switches to a front-facing stereo camera. It builds a 3D OctoMap in real-time for obstacle avoidance and employs a YOLOv3 model for door detection. If the door is not immediately visible, the drone follows a search pattern along the house footprint derived from the aerial segmentation.
Semantic Orientation and Navigation
Semantic orientation logic for marker-free navigation. Source, p. 4
Experimental Setup
The researchers utilized two primary datasets for training:
- Semantic Drone Dataset: Used to train a Unet architecture for aerial segmentation across 22 class labels.
- DoorDetect Dataset: Used to train the YOLOv3 model with over 1,200 images of doors and household objects.
- Evaluation: Testing was conducted in AirSim, a high-fidelity simulator, across 20 different house layouts to ensure robustness.
Key Results and Discussion
The proposed system demonstrated significant improvements in efficiency and reliability:
- Efficiency: The system reached the front door 161% faster than standard frontier exploration strategies.
- Accuracy: The semantic segmentation model achieved 83.6% overall accuracy, with high Intersection over Union (IoU) scores for roofs (0.81) and paved areas (0.76).
- Reliability: Drones successfully distinguished between front and back yards and navigated to the door while maintaining a safe 1-meter buffer for package release.
Conclusion
While successful in simulation, the system faces challenges such as misclassification in shadowed regions and sensitivity to initial GPS accuracy. Future work will focus on more intelligent search strategies for complex architectural designs and real-world field testing with human-provided instructions.
Summary updated