Autonomous Drone Delivery to Your Door and Yard

Shyam Sundar Kannan, Byung‐Cheol Min

TL;DR

Researchers at Purdue University have developed a marker-free autonomous system that allows delivery drones to find your front door or backyard just like a human courier would. By combining semantic image segmentation for high-altitude orientation and real-time visual navigation for the "last meter," the system reaches its destination 161% faster than traditional exploration-based methods.

What Problem Does This Paper Address?

While drone delivery is no longer science fiction, the "last-mile" (or rather, "last-meter") remains a hurdle. Current commercial solutions from giants like Amazon or UPS often rely on specialized landing pads, external markers (like QR codes), or lowering packages via tether to a general GPS coordinate.

However, most recipients prefer packages delivered specifically to their front door or a side yard—locations that GPS alone isn't precise enough to hit. Furthermore, requiring customers to place physical markers on their property is a barrier to mass adoption. This paper tackles the challenge of marker-free, semantic-aware delivery, where a drone understands the layout of a property (identifying the roof, driveway, and yard) to navigate safely to a specific user-requested spot.

Methods

The authors propose a two-stage pipeline to move the drone from a high-altitude GPS coordinate to a specific doorstep:

  1. Semantic Descending Location Estimation:
    • Using a downward-facing camera, the system performs semantic segmentation to identify features like roofs, paved areas (driveways/sidewalks), and grass.
  2. Orientation Logic: The system calculates the house's orientation by finding the boundary between the roof and the paved area. It then uses this vector to classify grass regions as either "front yard" or "back yard."
    • Safe Descent: It iteratively searches for a landing/descent spot that is at least 2.5 meters away from obstacles (trees, cars) and the house footprint.
  3. Vision-Based Navigation to Door:
    • Once the drone descends to a lower altitude (approx. 2m), it switches to a front-facing stereo camera.
  4. Mapping: It builds a 3D OctoMap in real-time for obstacle avoidance.
  5. Door Detection: The drone uses a YOLOv3 object detection model to find the door. If the door isn't immediately visible, the drone executes a search pattern along the house's footprint (derived from the earlier aerial segmentation) until the door is found.

Datasets

The researchers utilized two primary datasets to train their deep learning models:

  • Semantic Drone Dataset: Used to train the Unet architecture for aerial segmentation. It contains 400 images of urban environments with 22 class labels (roof, paved area, grass, etc.).
  • DoorDetect Dataset: Used to train the YOLOv3 model. It contains over 1,200 images of various doors and household objects.
  • Evaluation Environment: Testing was conducted in AirSim, a high-fidelity photo-realistic simulator, across 20 different house layouts to ensure robustness.

Key Insights and Ideas

  • Contextual Orientation: One of the cleverest "aha" moments is using the spatial relationship between the roof and the driveway to infer house orientation. Since driveways almost always lead to the front of a house, this allows the drone to distinguish between the front and back yards without needing a pre-loaded map.
  • Hybrid Mapping: The system uses a "top-down" occupancy map (from the aerial view) to plan the initial search path and a "front-on" 3D OctoMap to handle dynamic obstacles and fine-tune the approach.
  • Human-Centric Design: The system is designed to take natural language-style inputs (e.g., "deliver to the front door") rather than raw coordinates, making it more compatible with how logistics companies currently operate.

Key Results

  • Efficiency: The proposed system reached the front door 161% faster than a standard frontier exploration-based strategy (which explores the environment blindly until it finds the target).
  • Accuracy: The semantic segmentation model achieved an overall accuracy of 83.6%, with particularly high Intersection over Union (IoU) scores for roofs (0.81) and paved areas (0.76).
  • Reliability: In simulations across 20 different houses, the drone successfully distinguished between front and back yards and navigated to the door while maintaining a safe 1-meter buffer for package release.

Limitations and Future Work

The authors identify several areas for improvement:

  • Environmental Factors: The semantic classifier struggled with shadowed regions, sometimes misclassifying shadows on roofs as paved areas.
  • GPS Precision: If the initial GPS coordinate is too far off (closer to a neighbor's house than the recipient's), the drone might mistakenly identify the wrong roof as the target.
  • Complex Architecture: The current search strategy assumes a relatively standard house layout. Future work aims to implement more "intelligent" search strategies for complex architectural designs where the door might be hidden behind alcoves or fences.
  • Real-World Testing: The next step involves moving from the AirSim simulator to physical field tests with actual human-provided delivery instructions.

Summary updated

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.
BibTeX
@inproceedings{kannan2022autonomous,
  author = {Shyam Sundar Kannan and Byung‐Cheol Min},
  title = {{Autonomous Drone Delivery to Your Door and Yard}},
  booktitle = {2022},
  year = {2022},
  doi = {10.1109/icuas54217.2022.9836219},
}

Loading related content...

Figure