“The marine environment presents many complex conditions,” said Nina Mahmoudian, professor of mechanical engineering and principal investigator of the project. “Autonomous surface vehicles must deal with changing current or wind conditions. If there is a mechanical failure or if the cargo shifts, the system has to be resilient and adapt to those changing dynamics and maintain safe performance. So the goal is to find a new way to make autonomous boats smarter and more reliable in unpredictable conditions at sea.”

Purdue University has teamed up with Saab, Inc., manufacturer of a wide range of naval vehicles and systems. Together they are tackling this challenge from Defense Advanced Research Projects Agency (DARPA) as part of its Learning Introspective Control (LINC) program.

“Instead of relying on fixed models that can fail when the environment changes, we use a data-driven learning system that can adapt on the fly. ” Mahmoudian said. “We started with station-keeping. It sounds simple to keep a boat in one place, but it’s actually very challenging. Unlike a car that just sits on the ground, the maritime environment is constantly changing underneath the boat. Our approach combines advanced machine learning with safety features that detect when conditions shift and triggers the system to ‘relearn’ how best to hold its position.”

This simple boat, outfitted with multiple sensors, is a testbed for building machine-learning algorithms that will enable future marine vessels to function autonomously in changing conditions.

But “autonomy” doesn’t always rely on the machine itself. “What we are trying to do here is human-machine teaming,” said Christopher Vo, Chief Scientist for Autonomy and AI for Saab. “The idea is to have a human-in-the-loop, and that human is assisted by our algorithms at different levels.”

Sometimes, a vessel operator wants full control. Other times, some assistance from the algorithm would be helpful. And in some cases, the boat can dock itself autonomously.

“This is no different from the cruise control in your car,” Mahmoudian said. “Sometimes you want full control while driving, but the system alerts you if you need to brake quickly. Other times, you can put on adaptive cruise control, and it does most of the driving for you. Full self-driving is the final level of autonomy. With our algorithm, all of these can be accomplished on the water."

That requires a pretty robust algorithm, even in the face of challenging environments. “On the water, you have to deal with potential degradations or damage,” Vo said. “Anything can happen on the water, from engine cutouts to unforeseen obstacles. We test this by intentionally degrading the performance of one of the motors; but the algorithm adapts, and still successfully docks the boat.”

Proper autonomy functions on several different levels, from full self-driving to some degree of having a human-in-the-loop. The team worked just as much on the user interface as the algorithms.

From the lab to the lake

Mahmoudian’s team began in the lab. “First, we developed a high-fidelity computer simulation environment,” she said. “There we can develop our machine-learning algorithms and test them in virtual environments before we go out into the field. Our approach must compensate for unmodeled dynamics, payload variations, and hardware degradation while preventing unsafe behavior.” They enlisted the help of Inseok Hwang, Paul Stanley Professor of Aeronautics and Astronautics; Shaoshuai Mou, Elmer Bruhn Associate Professor of Aeronautics and Astronautics; and Shreyas Sundaram, Marie Gordon Professor of Electrical and Computer Engineering, for machine learning models and guaranteeing safe operation of the vehicle.

Of course, the real test is implementing them in the field. “For our test boat, we use a Wave Adaptive Modular Vehicle (WAM-V),” Mahmoudian said, “it’s outfitted with multiple sensors: LIDAR, cameras, wind gauges, inertial measurement units. It’s important to gather data about the boat as well as the surrounding environment, which can then be used to build our machine-learning models.”

Beyond just station-keeping, the team has also worked on autonomous docking — a safety-critical task for both manned and unmanned boats. “Autonomous docking is a much more complex challenge,” Mahmoudian said. “We developed a control system that helps autonomous boats dock safely, even in challenging conditions. It also includes a built-in safety feature that makes sure any steering or speed commands — even those given by a human operator — are automatically adjusted to avoid unsafe actions.”

Everyone involved in the collaboration has high aspirations. “Right now, we are working with a small boat in a small lake,” Mahmoudian said. “But these algorithms are scalable to bigger boats and ships. These types of systems will someday be used for search and rescue, for cargo transport, and for environmental monitoring. Learning how to transition these concepts from the lab, to the lake, and then beyond — that’s what makes this program very special.”

Team members from Purdue University and Saab are collaborating on the future of marine autonomy.

Distribution Statement A.  Approved for public release: distribution is unlimited.

Source: Nina Mahmoudian, ninam@purdue.edu

Writer: Jared Pike, jaredpike@purdue.edu, 765-496-0374