MA Thesis: Online Learning of Deep Neural Networks during UAV flight
Context
Deep Neural Networks (DNNs) have been applied successfully to mobile
robotics to solve challenging tasks in state estimation, path planning,
and control. Key disadvantages of DNNs are the need for large amounts of
labeled data for training and the task specific nature of their
performance. This thesis will aim to address these issues by exploring
online learning for DNNs during flight of the UAV.
Description
The thesis will build on existing work in the area of improved
propagation in an Extended Kalman Filter (EKF) through AI-based
pre-processing of high-rate inertial data (see
<https://doi.org/10.1109/ICRA46639.2022.9811989>). An initial
implementation of the online learning capable algorithm on the embedded
device of the UAV (Nvidia Orin) is available. The focus of the thesis
will be on the experimental work of training different DNNs by flying
the drone in the drone hall and evaluating their performance.
Tasks
- Analyzing and improving the existing code (C++, LibTorch)
- Conducting experiments in the drone hall
- Analyzing the performance of different DNN architectures and
hyperparameters
Milestones and Extensions
The milestones are an initial suggestion and subject to adaptation.
- M1: Code analysis and improvements finalized
- M2: Experiments for different DNNs and hyperparameters conducted
- M3: Performance of DNNs analyzed
- M4: Writing of Master thesis completed
Extensions
If successful, the student is encouraged and supported to submit their
results to one of the top international conferences in robotics.
Preferred Skill Set
- Good knowledge in training and evaluation of deep learning platforms
- Good knowledge Linux, C++, and Robot Operating System (ROS), or a
strong willingness to learn
- High level of self-motivation and responsibility
Period and Contacts
- Time period: 6 months, 1 student
- Skill board: Theory (1 of 5), Simulation (2 of 5), Implementation (5 of 5)
- Supervisors: Jan Steinbrener (jan.steinbrener@aau.at)
