Title of Project
Deep learning assisted underwater robot navigation
Name of Advisor/s
Daniel Cagara (AIMS), Paul Rigby (AIMS), Dmitry Konovalov (JCU)
Summary of Project
Marine robotics faces many unique challenges due to the complex and dynamic operating environment. The highly chaotic and cluttered nature of reef sites, the lack of a global positioning system underwater, reduced visibility, and imperfections in the obtained sensor data all present difficulties when designing navigation and control algorithms for autonomous underwater vehicles (AUVs).
With the recent advancements in machine learning, in particular large artificial neural networks with many layers ('deep learning') there is much potential to enhance the robot’s capabilities beyond those achievable with present methods.
The key idea of this project is to replace individual, “traditional” modules in robotics pipelines, with deep learning based modules in order to improve the platform’s performance and reliability. One could for example think of replacing traditional vision-based perception approaches    with a convolutional neural network approach that is able to “learn” to perceive the surrounding environment in varying conditions – even when the footage obtained from the sensors is highly degraded and would quickly bring traditional approaches to their limits.
Other parts of the robotics pipeline can profit from deep-learning based enhancements. Potentially this new technology could be applied to the navigation and control problem, and enhance the quality of the robots motion planning and decision making based on past experiences. It is even conceivable that the full end-to-end behavioural system could be replaced by a trained deep neural network with only raw sensor inputs, as recently demonstrated by Google DeepMind in .
The PhD student will, in the context of this project, have the opportunity to study deep learning technologies and explore how they can be adapted for the use in guidance and navigation of unmanned underwater vehicles to improve currently available capabilities and algorithms.
 Lindeberg, Tony. "Scale invariant feature transform." (2012): 10491.
 Bay, Herbert, Tinne Tuytelaars, and Luc Van Gool. "Surf: Speeded up robust features." European conference on computer vision. Springer, Berlin, Heidelberg, 2006.
 Mur-Artal, Raul, and Juan D. Tardós. "Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras." IEEE Transactions on Robotics 33.5 (2017): 1255-
 Mirowski, Piotr et. al. "Learning to Navigate in Cities Without a Map" arxiv:1804.00168,
Deep learning, artificial intelligence, neural networks, robotics, marine
Would suit an applicant who
The ideal candidate will have a strong academic background in computer science, mathematics, engineering or a closely-related numerical discipline. They will have strong mathematical skills and demonstrated proficiency with a programming language. Knowledge of artificial intelligence and/or robotics techniques is highly desirable.