The team tested the approach with 14 participants, each wearing a pair of ankle exoskeletons as they walked at a steady pace of about 2.3 miles per hour. The volunteers could take as much time as they wanted between choices, although they were limited to 50 choices. Most participants were choosing the same assistance profile repeatedly by the 45th decision.
After 50 rounds, the experimental team began testing the users to see whether the final assistance profile was truly the best — pairing it against 10 randomly generated (but plausible) profiles. On average, participants chose the settings suggested by the algorithm about nine out of 10 times, which highlights the accuracy of the proposed approach.
The control algorithm manages four exoskeleton settings: how much assistance to give (peak torque), how long to go between peaks (timing), and how the exoskeleton both ramps up and reduces the assistance on either side of each peak. This assistance approach is based on how our calf muscle adds force to propel us forward in each step.
Here is an exclusive Tech Briefs interview — edited for length and clarity — with Corresponding Author and Michigan Associate Professor Elliott Rouse.
Tech Briefs: What was the biggest technical challenge you faced while developing this technology?
Rouse: The biggest challenge was knowing if our algorithm or approach was successful. The idea behind this project was to receive feedback from the user that indicated what they enjoyed or what their preferences were around these different controller strategies. And it's really hard to know if we were successful.
So, the way we investigated that was by doing a blinded validation session where we stopped the algorithm from producing new options, and all we did was use the same option over and over again. That was the very last option, and we wanted to know if this was as close to their preference as we could get. And we found that, by using this blinded validation approach, we were able to get about 90 percent accuracy, which we felt very good about.
Tech Briefs: Can you explain in simple terms how it works?
Rouse: The game is to find the right assistance profile that feels good to the user and propels them forward and works with their existing neuromuscular control.
The way we're doing that is we have parameters that describe the curves. So, the game is to identify those parameters that maximize user preference. The way we chose to do that was by giving people two options, one of which they select. That selection is then used in an optimization approach called covariance matrix adaptation evolutionary strategy (CMA ES) — a computational optimization approach.
The idea is it's going to learn from the new settings that were selected and find a generation of candidates that it thinks, ‘OK, given what I know from your feedback, here is a generation of controller profiles that we think the algorithm thinks is going to be your favorite.’
Then we run that grouping or that generation through a neural network called RankNet, and it outputs what it thinks is the absolute best of that set of candidates; that best option is then provided back to the user via the touchscreen so they can try it out.
They’re trying out the new best vs. the old best. Then we take that feedback, update the model, run that through the CMA ES, and then find again the next best. And we kind of do this iteratively, and ideally what's happening is the person is providing feedback based on this touch interface of a pairwise comparison, and we learn from that to find the control or strategy that maximizes user preference and user enjoyment.
I think of it very similarly to Pandora radio. It’s like taking your thumbs up and your thumbs down and it's going to curate an assistance profile that maximizes your preferences.
Tech Briefs: First Author Ung Hee Lee is quoted as saying, ‘I'm excited this approach will make wearable robots … closer to becoming a normal part of our day-to-day life.’ How close to that ‘normal’ do you think we are?
Rouse: We’re much closer than people think! And I'd say what we're close to is going to Walmart and buying an exoskeleton that helps you walk. I think we are on the brink of seeing these technologies not necessarily in a medical application but in a recreational application.
Tech Briefs: What are your next steps?
Rouse: What I think we need to do next is enable these strategies to operate across all activities. So not just level-ground walking — a strategy that's not predicated on knowing the activity, but a strategy that's able to do non-cyclical on a voluntary motion without identifying a specific activity. That's the next major challenge.
Tech Briefs: Do you have any advice for engineers trying to bring their ideas to fruition?
Rouse: My advice is to be able to convert your effort into results. So, being able to accomplish the things that you want to accomplish and not spin your wheels. It’s really hard, especially for students, to be sure that they're making actual, tangible progress.
In terms of this field, I think people wanting to get involved in this field are going to have a lot of opportunity coming up as these technologies kind of exit the lab and make their way to everyday life. There are more and more companies starting up to build these technologies. So, I think there are going to be options — there are options now and there will be more options.