UiT Autonomous Ship Program
1 like109 views
This document discusses the UiT Autonomous Ship Program and its research on technologies to support autonomous maritime transportation systems. It proposes a ship intelligence framework (SIF) that uses deep neural networks (DNNs) trained on large datasets to mimic human ship navigator behavior. The goals are to overcome issues with ship controllability and replace human navigators. A decision support system would provide an adequate safety buffer to help DNNs handle unexpected situations. The framework is conceptualized based on factors behind successful self-driving cars, and aims to train DNNs using real-world ship navigation data to achieve accurate autonomous control.
UiT Autonomous Ship Program
- 1. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions UiT Autonomous Ship Program Current Status Lokukaluge P. Perera1, Peter Wide1, Bjørn-M. Batalden1, Ricardo Pascoal 1 & Brian Murray1 1Department of Technology and Safety UiT The Arctic University of Norway Tromso April 2019 UiT The Arctic University of Norway April 2019 1 / 23
- 2. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Outline 1 Introduction Transportation Systems Ship Maneuvering System Intelligence 2 Ship Intelligence Framework Introduction 3 Important Concepts 4 Navigation and Control Platform Ship Model Vessel Systems Onshore Command Center 5 Conclusions UiT The Arctic University of Norway April 2019 2 / 23
- 3. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Transportation Systems Autonomous Navigation Autonomous navigation will play an important role in future transportation systems. The technologies required for autonomous navigation in land transportation systems, i.e. self-driving cars such as Tesla, Uber, and Waymo, are in a mature phase when the environment is structured, i.e. well-defined roads and communication networks. The required technological advancements for autonomous transportation systems in an unstructured environment are subject to more challenging navigation constraints. Not only the required technologies for maritime transportation systems can be more complex and still in a development phase, the infrastructure is in general inadequate. A considerable amount of infrastructure and technology challenges has been encountered by maritime transportation systems in relation to autonomous navigation. This project proposes to research on the required fundamental technologies to support future maritime transportation systems operation under autonomous conditions. This requires an understating of the challenges associated with ship navigation and finding appropriate solutions. UiT The Arctic University of Norway April 2019 3 / 23
- 4. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Ship Maneuvering Vessel Controllability Problem Vessel position (i.e. the centre of vessel rotation) : P(x(t), y(t)) Course-speed vector : V(t) Heading (surge) vector : u(t) Sway velocity : v(t) Drift angle : β(t) Ship manoeuvring consists of complex rigid body motions. This is due to large bandwidth of nonlinear hydrodynamic and wind force and moment interactions between environment and the vessel hull and superstructure, which often generate unexpected and undesirable ship motions. Vessels often have a heading vector that deviates from the course-speed vector, resulting in a drift angle. Since vessels are not navigating in fully-defined ship routes, one vessel can encounter other vessels in its vicinity with various course-speed and heading vectors. Ship navigators should be aware of such encounters associated with higher collision risk. UiT The Arctic University of Norway April 2019 4 / 23
- 5. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Ship Maneuvering Vessel Controllability Problem (cont.) Vessel position (i.e. the centre of vessel rotation) : P(x(t), y(t)) Course-speed vector : V(t) Heading (surge) vector : u(t) Sway velocity : v(t) Drift angle : β(t) Full controllability of vessels with rudder and propeller actuators is not possible and is especially demanding under rough weather conditions. Vessels present sway-yaw manoeuvring interactions and are thus under-actuated systems with heavy inertia. The course-speed vector cannot be measured or estimated with enough accuracy, i.e. not enough sensor measurements. The centre of vessel rotation can also change due to the environmental loads. Not only the respective navigation vectors but also their positions can change without any requests. Vessels are considered as slow response systems with considerable time-delays in response to discrete control requests. UiT The Arctic University of Norway April 2019 5 / 23
- 6. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions System Intelligence Classical Mechanics 2 AI Various advanced controllers based on classical mechanics have been introduced by the research community to address this ship controllability problem , however the outcomes are still not satisfactory. The main reason is that these mathematical models may not adequately capture the complexities in ship motions; therefore controller robustness and/or stability cannot be preserved during ship manoeuvres. The controller inputs, i.e. rudder and propeller control inputs, are not continuous and the controller outputs, i.e. heading and course-speed vectors, may not have adequate accuracy and/or associated time-delays, hence controller performance can be further degraded. Though conventional ship auto-pilot systems are based on similar approaches, such systems may not able to handle complex navigational constraints. Ocean going vessels are still navigated by humans, especially in cluttered navigation zones and rough weather, using their knowledge and experiences to overcome complex vessel motions. Our aim is to overcome these issues in ship navigation by introducing Artificial Intelligence (AI) into the ship controllability problem. That has categorized as Cloning Ship Navigator Behavior. UiT The Arctic University of Norway April 2019 6 / 23
- 7. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions System Intelligence Self-driving Cars 2 Autonomous Ships This project proposes to capture and mimic navigator knowledge and experiences by using deep learning, i.e. deep neural networks (DNNs), as a ground-breaking technology that facilitates a better solution to control vessels. A considerable number of sensors should also be on-board and the respective data should be fused in a perception framework to achieve this objective. In self-driving cars, the respective driver is successfully replaced by DNNs trained to mimic human behaviour. The success in self-driving cars is due to three main factors 1 : 1 collecting and analysing large-scale real-world driving data sets, including sensor and high definition video/image data, to support deep learning based digital drivers. 2 holistic understanding of how human drivers interact with vehicle automation technologies by observing video/image and vehicle motion data, driving characteristics, human knowledge and experiences with the new technologies during the training phase. 3 adequate safety buffer to save lives by identifying how technology and other related factors can be used during the self-driving phase, i.e. the execution phase. 1. A. Fridman, et al., MIT Autonomous Vehicle Technology Study : Large-Scale Deep Learning Based Analysis of Driver Behavior and Interaction with Automation lessArXiv2017 UiT The Arctic University of Norway April 2019 7 / 23
- 8. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions System Intelligence Human-AI-Technology-Regulations Interactions UiT The Arctic University of Norway April 2019 8 / 23
- 9. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions System Intelligence Deep Neural Networks (DNNs) This is a considerable deviation from conventional control approaches in ship navigation developed in the last decade. Instead of control flow logics and if-then-else statements, DNNs consist of state-of-the-art Neural Networks with many layers and millions to one billion parameters, i.e. the weights of the respective neurons, of nonlinear activation functions. Convolutional neural networks are the most popular DNNs for self-driving cars and those network parameters are adjusted via back-propagation type approaches. DNNs require a large amount of real-world vessel navigation data and hundreds of thousands to millions of forward and backward training iterations to achieve higher accuracy in navigator behaviour. DNNs consist of large numbers of identical neurons with a highly parallel structure that can be mapped to GPUs (Graphics processing unit) naturally to obtain a higher computational speed when compared to CPUs based training. UiT The Arctic University of Norway April 2019 9 / 23
- 10. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions System Intelligence Adequate Safety Buffer It would be difficult build DNNs as a robust and safety critical system purely by human training. Unexpected and undesirable motion and navigation conditions can be encountered by ship navigation situations. If the DNNs have not seen such situations during its training phase and generalization is poor in the execution phase, that could create undesirable behavior. A decision support layer with adequate information sources should support the DNNs to overcome such situations. Situation awareness and collision avoidance (SACA) is identified as the minimal decision support facility required to support the training and execution phases. This can create an adequate safety buffer to avoid possible collision or near-miss situations, by identifying moving and stationary objects around the vessel domain. DNNs are integrated into the ship intelligence framework (SIF) created in a conceptual level by the UiT autonomous ship program, while considering the same main factors of self-driving cars. UiT The Arctic University of Norway April 2019 10 / 23
- 11. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Introduction SIF (cont.) UiT The Arctic University of Norway April 2019 11 / 23
- 12. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Introduction SIF in The Training Phase UiT The Arctic University of Norway April 2019 12 / 23
- 13. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Introduction SIF in The Execution Phase UiT The Arctic University of Norway April 2019 13 / 23
- 14. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Introduction A View from the Bridge Information Visualization Platform (IVP) on the bridge. Decision support system for ship navigators during the training phase and DNNs during the execution phase. Required ship route as the Digital Ship Route (DSR) Actual & Predicted ship route as the Advanced Ship Predictor (ASP). Local scale with ship performance and navigation data. Global scale with AIS data. The Situation Awareness and Collision Avoidance (SACA) Module. Target Detection and Tacking Unit (TDTU) Collision Risk Assessment Unit (CRAU) UiT The Arctic University of Norway April 2019 14 / 23
- 15. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Advanced Ship Predictor FIGURE – Local scale FIGURE – Global scale UiT The Arctic University of Norway April 2019 15 / 23
- 16. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Situation Awareness and Collision Avoidance (SACA) FIGURE – Local scale FIGURE – Global scale UiT The Arctic University of Norway April 2019 16 / 23
- 17. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Ship Model Autonomous Vessel Small scale vessel will be purchased. UiT The Arctic University of Norway April 2019 17 / 23
- 18. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Vessel Systems GNSS & INS System UiT The Arctic University of Norway April 2019 18 / 23
- 19. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Vessel Systems Engine, Rudder & Propulsion System UiT The Arctic University of Norway April 2019 19 / 23
- 20. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Vessel Systems Other Sensors UiT The Arctic University of Norway April 2019 20 / 23
- 21. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Vessel Systems Communication System UiT The Arctic University of Norway April 2019 21 / 23
- 22. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Onshore Command Center Initial Experiments Ferry crossing type maneuvers will be conducted under this vessel, initially. Autonomous Test-Site in Tromso will be developed with the legal requirements. Ashore Remote-controlled Center will be developed near the test-site. Ashore Operational Center will be developed in UiT to visualize sensor data for teaching and research purpose. UiT The Arctic University of Norway April 2019 22 / 23
- 23. Introduction Ship Intelligence Framework Important Concepts Navigation and Control Platform Conclusions Any Question? UiT The Arctic University of Norway April 2019 23 / 23