On 19 September 2020, to commemorate the 400th anniversary of the Mayflower’s voyage to the New World, an autonomous trimaran vessel, the Mayflower Autonomous Ship, will trace the route of the original Mayflower in 1620, sailing from Plymouth, UK, to Plymouth, Massachusetts, US.
Although remotely controlled ships are not a new concept, Don Scott, chief technology officer of the Mayflower Autonomous Ship, says the project is at the bleeding edge. “What’s new about the project is the marine autonomy aspect, creating the ship as an edge device that operates on its own, sensing its environment, making intelligent decisions and acting on them without any human intervention,” he says. “That’s what makes this vehicle really innovative.”
The Mayflower is a prototype, a proof of concept, to demonstrate that a solar-powered autonomous ship can navigate the oceans safely and cope with changeable weather, other ships, and encounters with marine creatures and submerged hazards.
Scott has worked on the oceans for 30 years. For him, one of the biggest challenges in developing an autonomous ship is the unpredictable nature of the sea. “You certainly don’t approach these engineering tasks lightly,” he says. “You very quickly get humbled by the power of the ocean.”
He says one of the philosophies that underpins the design is that the Mayflower Autonomous Ship needs to operate in an extremely hostile and dynamic environment, which is very unpredictable.
Until very recently, undertaking such an engineering task would have seemed impossible. For Scott, the recent convergence of technologies such as computing power at the edge and the growth of machine learning has meant that today it is possible to have an edge device make decisions in a timeframe that enables a vehicle to operate within the environment it was designed for.
Engineering for unpredictability requires partitioning different tasks, so that there is a strong sense of separation and layering between the software running on the edge devices. The architecture is highly modularised, where each edge device maintains its own situational awareness and communicates upstream and downstream with other modules.
Sensor inputs include six cameras, an automatic identification system, wind speed and direction sensors and obstacle avoidance sonar. Scott says: “Each of these collects unstructured data, which is then processed and fed into a data server to provide the information needed for the vessel to navigate.”
Weather forecasts are provided via application programming interfaces (APIs) to The Weather Company. “Weather updates will be our highest priority,” says Scott. “We will steer around a storm cell, for sure.”
The Mayflower is designed to run autonomously, but it will have the ability to send and receive data. Given that the ocean offers limited low-bandwidth satellite communications, weather data is given a priority. “It is critical information – we will get what we can get,” says Scott.
IBM PowerAI Vision models are being used to provide object classification and object tracking for the vessel’s computer vision system, he says, adding: “All of this information is fed into a navigation hazard map used by a collision avoidance module.”
The collision avoidance module takes this data to determine a series of potential courses and speeds, which are then fed into a route planner, which Scott says is essentially an autonomous system that determines the course the vessel should take. “Layered on top of the route planner is our safety manager, which deals with more localised information, such as wave direction and the unpredictable aspects of the ocean,” he adds.
Each system on the vessel is redundant, with a backup module running in parallel, which is ready to take over if the primary system fails. One of the hazards of the ocean is short circuits, so the Mayflower Autonomous Ship has been built in a way that enables it to continue if systems are damaged.
“Any system that is exposed to the ocean needs to be isolated to protect against electricity shorts,” says Scott. “We are hedging our bets on the electrical reliability of the system by putting in a backup system.”
Because each system has been designed to operate independently, each one can be tested before they are all integrated on the ship, says Scott. For instance, at the start of March, the collision avoidance system, called AI Captain, is being tested at sea on another ship.
“The ocean presents a lot of different hazards, such as land, marine debris, submerged objects, wildlife and even curious whales,” says Scott. “In a classical marine system, human vigilance is required to make decisions on these hazards.”
The autonomous system also needs to adhere to the rules of the sea, he says. “We need to identify surrounding marine traffic and make sure we are operating safely.”
Scott says IBM suggested taking ODM, its rules-based engine for determining credit card fraud, and adapting it to marine regulations. This means that AI Captain enables the Mayflower to follow the International Regulations for Preventing Collisions at Sea (COLREGs) as well as recommendations from the International Convention for the Safety of Life at Sea (SOLAS).
Having the AI Captain make decisions based on a predefined set of regulations means its decisions do not come out of a black box, says Scott. In effect, the AI decision-making for collision detection is entirely explainable because it adheres to the rules that govern marine navigation.
For Scott, the vehicle systems the Mayflower will use are proven technologies, tried and trusted. “We know we can do this voyage tomorrow, with the existing capabilities we have in classic deterministic systems, going waypoint to waypoint, and dynamic updates based on local conditions,” he says.
But the fully autonomous AI Captain needs to operate without any human intervention. Scott adds: “The thing that keeps me up at night is the collision avoidance system, which we haven’t tested yet. It is essential for the success of the ship. We need to go through a bunch of sea trials.”
These trials are beginning in early March.