Intelligent trajectory predictions at sea using neural ordinary differential equations

Industrial PhD student project in collaboration with Saab Kockums within the Wallenberg AI, Autonomous Systems and Software Program (WASP), funded by the Knut and Alice Wallenberg Foundation.

Project outline and scientific objectives:
This project develops methods for trajectory predictions of surrounding vessels at sea, by modeling of the interacting vessels combined with neural ordinary differential equations (ODEs), relying on input from a perception system, with the purpose of presentation to a human operator or utilization in an automated planning and control system. Operators of sea vessels make decisions about actions based on the current situational understanding, and how they predict various types of vessels will behave in the specific environment that they are operating in. Having models for such functionality, is a requirement for high-performance (semi-)autonomous sea vessels, and is significantly relevant for both manned vessels and autonomous surface vessels in dense and narrow environments like archipelagos and ports. Compared to the established approach in use in many systems, where trajectories of surrounding vessels are extrapolated holding a straight line and a constant velocity, the approach taken in this project contributes to a more realistic situational understanding and prediction by relying on an interaction-aware and physics-aware model that ensures physically reasonable and situation-specific trajectory predictions as output from the model.

Research questions:
The research questions that are the focus of this project are:
- How can modeling of interactions combined with neural ODEs and other known information be used for model-based trajectory predictions to achieve robust and safe naval autonomy?
- How should uncertainty for trajectory predictions be quantified and actively managed by interactions with the perception system in multi-agent settings for sea vessels?
- How should the system architecture for implementation of model-based trajectory prediction systems be designed to enable interactions with surrounding system components for full autonomy in naval applications?