If humans walked like robots, engineers already would have perfected zero-effort, mechanically-assisted walking. But what about people who bounce on their toes, power walkers, those who sashay? Habits, diseases, and disabilities can affect someone’s gait in unique ways. An idealized exoskeleton needs to be both easily accessible and personalized.
The Chipotle of exoskeletons doesn’t quite exist yet. Computers still struggle to anticipate how people will move—they’re literally a moving target. From a data standpoint, humans are noisy, says Katherine Poggensee, a biomechatronics researcher. Plus, “they have brains, so they adapt over time.” And although humans generally find the easiest way to do any motion, very few people have the physical and spatial awareness to explain why one stride feels easier than another. That’s why researchers are turning to algorithms to make exoskeletons more efficient.
So far, automatically tuning an exoskeleton’s force, and the timing of that oomf, is faster and better than hand-tuning. Thursday, in a paper published in Science, Poggensee and her fellow researchers outline an algorithm that calibrates an exoskeleton to best assist its user. To do that, they use a type of optimization that’s also helped govern how animated characters interact with their environments in CGI.
Instead of supplying users with standardized assistance, these control algorithms set themselves up like an eye doctor who flips through lenses while asking “better, or worse?” But instead of actually asking users, the algorithms rely on sensor feedback. To minimize the energy required to walk, for example, they track respiration to calculate metabolic rate, then optimize to minimize the calorie burn.
This algorithmic tuning can only happen in a lab, on a treadmill, where there are machines to perform and analyze these extra measurements. The idea is that eventually, you could get fitted for your exoskeleton or robotic prosthetic limb in a clinic, then transfer your personalized profile to the outside world. And in this study as well as others, automatically tuned exoskeletons do successfully lower the energy it takes to walk.
This is an improvement over previous versions of exoskeletal tuning, which were slower, and in some cases, demanded more effort than normal non-assisted walking. For simpler approaches that relied on a brute-force sweep through many different options, “the numbers get really hard to deal with,” says Daniel Ferris, who has developed similar algorithms to calibrate exoskeletons. There are different mathematical approaches to automating this tuning, but the most effective ones all start by guessing how a human will respond, then monitoring their actual response while offering up different calibrations.
Because the algorithms also incorporate stochasticity, or randomness, into their structure, the exoskeletal controllers evolve differently for each walker. In the method published this week, the controller starts off by trying eight different tuning profiles. Based on which of those work well, it generates eight new profiles to try, with a few wildcards thrown in. Sometimes the wildcards are better, and other times worse, but they all force the controller to evolve. As the wearer inevitably adapts to the exoskeleton’s assistance, the control loop also adapts to the wearer.
For Poggensee’s proof-of-concept tests, 11 human guinea pigs donned an ankle exoskeleton over one of their shoes and took a stroll on a treadmill. As they walked, a respiratory mask measured the oxygen they inhaled and the carbon dioxide they exhaled, calculating the energy cost of walking. Meanwhile, the tuning algorithm cycled through four sets of eight different patterns of assistive torque, varied in timing and amount of force.
After about an hour of this strolling, the algorithm pinned down the optimal timing and torque to minimize the energy cost of each walker’s gait. Each participant’s ideal pattern was different—a little more help at toe-off, less force at the middle of the stride–so that when you look at the torque profiles of all the walkers, you see “a bunch of different shapes,” says Poggensee.
Energy expenditure, of course, is only only one way to assess the effectiveness of an exoskeleton. Studies like this one can also quantify activity by monitoring voltage across local muscles, using a method called electromyography. But there are plenty of other metrics to optimize, like heart rate, limb speed, and balance. Or, if you’re willing to delve into the wild west of subjectivity, comfort and perceived effort.
Taking those additional factors into account—and expanding those factors to address a wider range of needs—could be more of a challenge, says Ferris. He points out that these optimization methods do well with a handful of parameters in the lab, but the real world ultimately demands control of many knobs at nearly infinite settings. Navigating a crowded subway car, for example, requires attention to more than just energy. There’s also minimization of exposure to armpits, and additional calibration for manspreading. Before those factors can be optimized, they’d need to be measured—which might be work for another algorithm entirely.