An ML Algorithm Which Copies the Wearer's Knee Could Improve Future Robotic Prostheses
Machine learning system offers accuracy between 97-100 percent, depending on terrain, for steady-pace walking.
A team of researchers working at Arizona State University and North Carolina State University have figured out a way to have robotic leg prostheses and exoskeletons behave more realistically — by using an intact knee, and a reinforcement learning algorithm, as an "instructor."
“Classical control is a vast field with elegant results, but those methods usually require some knowledge of how the human and robot mathematically work together," explains Jennie Si, processor, of her tea's work. "This is difficult to do because people are so different — some are short or tall, some walk fast, some lift their feet higher, so we walk in different patterns. That’s much harder for a robot to understand."
Rather than creating a mathematical model of a knee, then, Si and team took a different approach: A reinforcement learning algorithm, which took its cues from the user's intact knee, copying its movement with the robotic version, dubbed direct heuristic dynamic programming (dHDP).
The results were impressive: Based on walking on even terrain, the robotic knee copied its natural twin with 100 percent success; shifting to uneven terrain, this dropped to 97 percent. Only when the user began to vary their pace, speeding up and slowing down unpredictably, did the accuracy drop to 80 percent — but this came back up with additional training.
"This is a dynamic problem, meaning the variables evolve over time," Si explains. "The prosthetic can work in symmetry with the human to create a comfortable experience. The user doesn't need to stop and say, 'I am going up stairs now,' and wait for the robot to figure out how to do that. They can just go up the stairs together.
"People have only recently realized the potential of machine learning to improve our lives. With this type of artificial intelligence, we can solve problems to make people's lives better."
The team's work has been published in the IEEE/CAA Journal of Automatica Sinica under open-access terms.