General Goal:
In the worldwide R&D on computer-assisted and autonomous navigation the DDSHIP project will contribute by setting a new process flow methodology and test platform for validation and certification through investigations on:
·       more accurate and robust perception and situational awareness of the waterborne world around the ship in dense traffic and harsh weather conditions;
·       the accurate representation of the real behaviour of the ship in complex waterways with low under keel clearances and nearby banks and infrastructure;
·       the safe and smooth control of the ship through model predictive AI-trained controllers providing necessary collision avoidance.
As accidents on waterways are mainly attributed to human actions in combination with failures of technical hard- and software or environmental circumstances, the support of captains, pilots or skippers on board the manned ship or the operator from a remote operation centre on an unmanned ship, this research should prove the capabilities of existing technologies (camera, sensors, manoeuvring model prediction, path-planning and steering) leading to smarter - more accurate and higher reliability – control.
Concrete objectives and criteria
The general goal in the realization of an accurate data-driven autonomous ship is related to three knowledge domains of (1) sensors and sensor fusion, (2) ship manoeuvring in shallow and confined water and (3) deep reinforcement learning with the following objectives and criteria:
Data and sensors:
·       Using a multi sensor 3D vision data simulator, we will investigate the robustness of different camera technologies (in the visible and infrared wavelength range) for automated ship detection under different environmental conditions (fog, rain, sunlight). The detection uncertainties that we will obtain from the simulations will allow us to optimally combine different sensor types in harsh environments. Moreover, we will be able to use these uncertainties as inputs for the identification and control process of the ship behaviour.
·       The amount of available data has to be investigated so that impact on the controlling in the final step is proven throughout the complete data pipeline from rough data to precise control.
Domain knowledge:
·       The use of in depth domain specific knowledge in the different intelligent components:
o   Ship modelling: The so-called control plant (the ship) is affected by the environment . In short the environment[1] changes the flow around the ship, creating additional hydrodynamic forces. As a result the ship behaves differently than what a human/artificial operator would expect if he faces this issue for the first time (in other words if neither received appropriate training in this new environment). The relationship between appropriate control actions and the hydrodynamic behaviour is thus not sufficiently well understood.
o   Model predictive reinforcement modelling (including the model of the ship and the environment in the Markov decision process) with performance investigation of different model types: complex or simplified physics-based models or data-driven surrogate models.
Control strategies:
·       The impact of environmental circumstances through the control pipeline from perception to control.
·       End-to-end training (perception <-> control) instead of component based training. The current state-of-the-art trains component by component (perception of the plant model/ perception of the environment model/control model). With end-to-end training the risk of losing important data due to incomplete representations between components will be reduced. This is particularly important in harsh situations where irrelevant data for perception in low quality might make a difference in the control phase.
·       Thanks to the control based on simplified modelling or high accuracy manoeuvring models the end- to-end training can be brought to higher levels.