Simulations and ML for parachute guidancePublic
Downloadable Contentopen in viewer
In the military, supplies are critical, and a common method for supply deliveries are parafoil parachutes. The accuracy of these supply drops is important to the success and safety of operations and supporting American troops in their missions. The most common way of tracking and navigating these parafoils is via Global Positioning System (GPS). However, signals from GPS satellites can suffer from disruptions due to interference like jamming or environmental factors. The drawbacks of GPS signaling have led to research in alternative forms of navigation that do not rely on external signals. One in particular is the area of Guided-Airdrop Vision-Based Navigation (GAVN). In recent years there has been an increase in machine learning (ML) and Data Science (DS) research. Some of the most popular ML models for digital image processing are neural networks (NN), a type of deep learning model that can be trained using examples in order to make predictions on new data. Deep learning algorithms have the potential to pinpoint the location of cargo parafoils by using the visual feed from a camera mounted to the parachute, as opposed to external GPS signaling. Accordingly, the research in this MQP focuses on the development of a machine learning-based alternative to GPS navigation for cargo parafoils. In order for neural networks to make meaningful predictions, they require large amounts of training data. However, collecting images from parafoil drops is expensive and resource-intensive. Therefore, we first created a virtual simulator for parachute drops. This simulator is able to generate the large quantities of labeled image data required by our proposed deep learning methods in an inexpensive manner. Next, after collecting data using our simulator, we were able to apply a variety of preprocessing methods to the images and test different neural network structures in an attempt to predict location from the images. In particular, we studied the effect of multi-task deep learning models on the performance of vision-based navigation in realistic parafoil scenarios. Additionally, we found through our experiments that multi-task learning yield more consistent error results compared to single-task networks for our vision-based navigation models.
- This report represents the work of one or more WPI undergraduate students submitted to the faculty as evidence of completion of a degree requirement. WPI routinely publishes these reports on its website without editorial or peer review.
- UN Sustainable Development Goals
- Date created
- Resource type
- Rights statement
- In Collection:
Permanent link to this page: https://digital.wpi.edu/show/sj1395264