Trajectory Estimation of Unmanned Vehicles With Model-Based and Learning-Based Algorithms

The goal of this project is to design and analyze algorithms that estimate the state of a drone (e.g., its position and velocity) from partial or compressed information. The student will develop, study, and implement new algorithms based on neural networks, nonconvex optimization, and control theory. The student will also conduct an analysis of the fundamental tradeoffs between estimation, compression, and target dynamics.
Description of the Project: 

Drones play an increasingly important role in commercial and military applications. For example, Autonomous Underwater Vehicles (AUVs) are essential to make detailed maps of the sea floor, to inspect bridges, to survey marine life, and to track ocean pollution. The underwater environment, however, is one of the worst environments to communicate information in, either among AUVs, or between AUVs and a base station. The main reason is absorption loss and multipath reflection of electromagnetic and acoustic waves. As a result, underwater communication has low bandwidth and high latency, requiring efficient information compression strategies. Another example is aerial drones, also known as Unmanned Aerial Vehicles (UAVs). UAVs will be increasingly used for agriculture, package deliveries, search and rescue operations, combating wildfires, or building inspections. They can, however, also be used with malicious purposes, including disrupting airports or transporting dangerous payloads to specific targets. To counteract these adversarial cases, it is essential to accurately estimate the position and velocity of the UAV from limited distance measurements (LiDAR), remote RGB cameras, or both. 

Both AUVs and UAVs move according to specific models, which may be fully known as in the case of underwater exploration, or partially known as in the case of tracking an adversarial UAV. Conventional state estimation methods are based on Bayesian techniques, which have limited computational performance and tend to rely strongly on statistical assumptions about the problem.  

In this project, we draw from recent developments in neural networks, and nonconvex optimization, to design and analyze algorithms that estimate the state of a drone, described via a state-space model. In the case of AUVs, we also seek to design algorithms that efficiently encode and decode information about the state. The main challenge of designing neural networks to perform these tasks is that it is not obvious how to integrate into their design the information of how the drone moves, that is, the known or partially known model. To achieve this, we will leverage recent techniques that enforce model consistency in image processing tasks. In the project, we will also seek to study information-theoretic tradeoffs between the smallest possible sampling rate, the state compression level, the dynamics of the vehicle, and the smallest achievable estimation error. 

As the project is in partnership with SeeByte Ltd., there is a great opportunity to implement the developed algorithms in bespoke AUV simulators, or even in real AUVs. 

 

Project number: 
140031
First Supervisor: 
University: 
Heriot-Watt University
First supervisor university: 
Heriot-Watt University
Essential skills and knowledge: 
This project requires a strong mathematical background (control theory, optimization, analysis of algorithms, information theory) and also some practical skills, including expertise in Matlab/Python/Julia
Desirable skills and knowledge: 
Deep learning Compression techniques C/C++, ROS