A new artificial intelligence controller for exoskeletons that can learn various human movements without requiring special programming marks a major step forward in wearable robotics by saving significant amounts of energy.

Imagine safer, more efficient travel for factory workers and astronauts, and improved mobility for people with disabilities. This may one day become a more common reality, thanks to a new study published June 12 in the journal Nature. Nature.

The first author of the article is Dr. D. from Embry-Riddle Aeronautical University. Shuzhen Luo explained that wearable robotic frames for the human body, called “exoskeletons,” promise easier movement, but technological hurdles limit their broader use. Nature Corresponding author Dr. With Hao. Su and colleagues from North Carolina State University (NC).

Currently, exoskeletons must be pre-programmed for specific activities and individuals based on lengthy, expensive and time-consuming human testing, Luo noted.

Introduction to control with artificial intelligence

Now researchers have identified a super-smart, or “learned,” controller that uses artificial intelligence (AI) and data-intensive computer simulations to train portable robotic exoskeletons.

“This new controller provides seamless, continuous torque support without the need for human testing when walking, running or climbing stairs,” Luo said. “With just one run on the GPU, we can teach a control law or ‘policy’ in the simulation, so the controller can effectively assist three activities and different people.”

Revolutionary energy reduction

D., an associate professor of mechanical engineering at Daytona Beach Embry-Riddle, Florida. Guided by three interconnected multilayer neural networks, the controller learns as it goes and “evolves through millions of periods of musculoskeletal simulation to improve human mobility,” Luo said. . campus

The experiment-free “learning in simulation” framework applied to a custom hip exoskeleton resulted in the highest reductions in metabolic rate of portable hip exoskeletons to date, averaging 24.3%, 13.1%, and 15.4%. % reduction in the energy spent by users while walking, running and climbing stairs, respectively.

NC State’s Su explained that these energy reduction rates were calculated by comparing the performance of humans with and without robotic exoskeletons. “This means it’s a true measure of how much energy the exoskeleton stores,” said Su, an associate professor in the Department of Mechanical and Aerospace Engineering. “This work actually turns science fiction into reality, allowing people to use less energy when performing a variety of tasks.”

Bridging the gap between simulation and reality

The approach is believed to be the first to demonstrate the feasibility of designing controllers in simulation that overcomes the so-called simulation-reality gap, or “sim2real gap,” while significantly improving human performance.

“Previous advances in reinforcement learning generally focused on simulations and board games,” Luo said. “Whereas, we proposed a new method; a data-driven, dynamic reinforcement learning method for training and controlling wearable robots that will directly benefit humans.” .

Su added that the framework “can offer a generalizable and scalable strategy for rapid and widespread deployment of various assistive robots for both healthy humans and people with limited mobility.”

Overcoming technological barriers

As noted, exoskeletons have traditionally required hand-crafted control laws based on labor-intensive human testing to perform each activity and account for individual gait differences, the researchers explain in the journal Nature. The simulation learning approach offered a possible solution to these obstacles.

The resulting “dynamics-based and data-driven reinforcement learning approach” has greatly accelerated the development of exoskeletons for real-world applications, Luo said. Closed-loop simulation includes both the exoskeleton controller and physical models of musculoskeletal dynamics, human-robot interaction, and muscle responses to generate effective and realistic data. In this way, the control policy can be developed or learned in simulation.

“Our method provides a framework for off-the-shelf solutions in the design of controllers for wearable robots,” Luo said. said.

Future directions of exoskeleton research

Future research will focus on a unique gait for walking, running or climbing stairs to help people with disabilities such as stroke, osteoarthritis and cerebral palsy, as well as amputees.