Photo credit: Kevin Murrant, NRC
Scientists are often inspired by nature. For some years now, mathematicians, computer scientists, robotics specialists and others have explored and experimented with the idea of swarm intelligence, drawing inspiration from bees, ants and other animals that have an apparent ability to work together towards a common goal—such as building a hive or nest—without centralized, top-down instructions.
Now, researchers and engineers at the National Research Council of Canada (NRC) and Memorial University of Newfoundland in St. John's are applying this idea to the challenge of navigating in ice-covered seas. Their "swarm" is a fleet of self-driving, uncrewed vessels. The collective goal for these autonomous vessels is to clear sea ice from Northern shipping channels. The team is studying ways of controlling the fleet to achieve that goal through an artificial intelligence (AI) technique known as reinforcement learning.
An icebreaker's work is never done
To find their way through ice-laden waters, ships often rely on an icebreaker, a robust vessel designed not only to withstand ice but also to break it apart. The icebreaker gets the job done, but in harsh conditions, the channel it clears can quickly close over again. Dr. Kevin Murrant, an engineer with the NRC's Ocean, Coastal and River Engineering Research Centre, and Dr. Andrew Vardy, who runs the bio-inspired robotics group (BOTS) at Memorial University, teamed up to take a fresh look at this problem. The question underlying their approach: could a fleet of smaller, self-driving vessels come in after an icebreaker and push the ice out of the way to keep the channel clear indefinitely?
A novel solution for a complex problem
For the most part, seafarers want to avoid hitting ice. But with their system, Murrant and Vardy are aiming to do the exact opposite—steer them towards the ice to move it out of the way—and that's a goal that comes with challenges of its own. As Dr. Murrant explains, ice does not stay in one place on the sea surface, which makes it especially difficult "to arrange the ice field into a certain configuration, for example, to push all the ice to the side of a channel." Add to that the idea of a number of uncrewed vessels working together and the problem gets even more interesting.
Dr. Vardy describes the swarm approach as "a collaboration of agents that does not rely on a central intelligence." One of its main advantages, he says, is its scalability. If one vessel breaks down, others can take over or more vessels can be added as needed.
Dr. Vardy adds that the way ships operate and the conditions in which they operate make the task of creating an AI system to control them complex. "Ships are unique systems because they are slow acting. They take a lot of time to come to a stop. Ship interactions are complicated by waves, and because ships also create waves."
Reward-driven machine learning
The team says reinforcement learning is well suited to this problem because it typically requires a clearly defined end goal. In this case, the goal is that an area becomes ice-free.
Dr. Marius Seidl is a member of the BOTS group and working towards his second PhD. He explains that, for reinforcement learning to succeed, one needs to define a goal and a "digital reward" for getting closer to the goal. "If the uncrewed vessels move the ice out of the way, the system gives a high reward. However, if the surface area covered by ice increases, it gives a negative reward."
The researchers define what the final state should look like, and the AI learns how to achieve that. Dr. Seidl compares it to ants forming a colony. "No single ant knows the entire plan." Yet, they succeed in building intricate nests. That type of decentralized collaboration is what the researchers are trying to recreate.
Testing the theory with digital and physical modelling
Dr. Seidl has designed a simulator, essentially a computer model, to test the team's AI. Once they're satisfied that their swarm can tackle the problem in this simulated environment, they will take it to the "real world" of the NRC's offshore engineering basin research facility. Here they will test the performance of model vessels operating among ice floes—some made of plastic that matches the density of ice, others from sea ice created in the lab.
"Moving ice around is important for the shipping and oil and gas industries, which are both economically important in Newfoundland and Labrador," says Dr. Vardy, noting the ferry that provides a vital link between Labrador and the island of Newfoundland across the Strait of Belle Isle is routinely accompanied by an icebreaker in the winter.
Dr. Vardy sees the potential for their swarm ice removal service to assist in the work of manually operated icebreakers. For instance, a sheet of ice blocking a passage or harbour could be opened up by an icebreaker and then a group of smaller self-driving vessels could work to keep the area clear of ice.
Positioning Canada as a leader in the field
The NRC's support for this project is one example of how we're working to develop Canadian leadership in the rapidly growing field of marine autonomous surface ships, or MASS.
"It's great to be working on something new and in vogue in the international landscape. We have this opportunity to be pioneers in the development of MASS," Dr. Murrant says.
"We think the best way Canada can leverage that is through our harsh environments," he adds, referring not only to working with ice but also to Canada's long coastline and challenging marine weather and wave conditions.
"It's worthwhile doing something that doesn't yield a great solution tomorrow, going to this uncharted territory with a big new idea and trying to get something going. I enjoy playing with robots and solving puzzles. And that's what research is. Every day is different. We're facing problems and finding solutions, discovering new knowledge that might make the world a better place 20 years down the line."
Reinforcement learning is a form of artificial intelligence that relies on trial and error to achieve a predefined goal. In a reinforcement learning model, an agent is rewarded for the attempts it makes at achieving a goal: the closer it gets, the higher the reward. The agent learns to make decisions to maximize the reward and, over time, discovers increasingly effective ways to achieve the goal. This type of AI is ideally suited to solving well-defined problems like clearing pieces of ice from a harbour.
Contact us
Media Relations, National Research Council of Canada
1-855-282-1637 (toll‑free in Canada only)
1-613-991-1431 (elsewhere in North America)
001-613-991-1431 (international)
media@nrc-cnrc.gc.ca
Follow us on X: @NRC_CNRC