For those with physical disabilities and workers in physically demanding jobs, exoskeletons can enhance mobility, increase strength, and help to prevent injury. Or at least they could if anyone was actually able to get their hands on one. These systems are very rarely used today due to their complexity and high costs.
The main source of the complexity and costs may not be what you think it is. The hardware itself is the most obvious culprit, but the controller is often more difficult and expensive to prepare for each individual user. The controller is an essential piece of the overall system; it determines exactly when, how, and how much to assist the wearer. This assistance must be carefully tuned to work with the unique body mechanics and capabilities of each exoskeleton wearer. Without a properly functioning controller the exoskeleton could potentially harm the wearer, so there is no room for skimping in this area.
Fortunately, there may be a way to make exoskeleton technology more accessible. Engineers at Georgia Tech have developed a new type of controller that takes a one-size-fits-all approach to the problem. Rather than explicitly training it to work with a single user, it was trained to understand the needs of many people. This allows it to recognize what type of help its wearer needs and respond accordingly.
Members of the research team working on their system (📷: Georgia Tech)
Traditionally, developing these controllers has required enormous amounts of time, money, and human data collection. Each new exoskeleton model forced researchers to start from scratch, gathering hours of movement data from the people using the device. This expensive development cycle has kept exoskeletons largely confined to research labs.
But by leveraging an AI model known as a CycleGAN, originally used to map satellite images to ground-level views or transform horses into zebras, the researchers discovered a way to convert massive datasets of how people move without an exoskeleton into predictions of how they would move with one. In other words, the AI acts as a translator: it takes one “language” of human motion and transforms it so any specific robot can understand what kind of help a person will need.
This translation ability means exoskeleton developers no longer need to collect fresh data each time a device changes. A startup could iterate through several hardware designs without retraining the controller from scratch, slashing both development time and cost.
Using this approach, a controller can provide meaningful assistance across a wide range of hip and knee movements. It doesn’t attempt to guess a user’s intention — such as whether they are climbing stairs or stepping off a curb. Instead, it analyzes joint motion in real time to estimate how much effort the user is exerting and boosts that effort by up to 20 percent. Tests with a lower-limb exoskeleton showed that the new approach performs on par with the best specialized controllers developed through years of traditional data collection.
Although the study focused on leg exoskeletons the same AI-driven system could potentially be applied to upper-limb devices, prosthetics, and autonomous robots. The team hopes that by lowering the barriers to controller development, more engineers will be able to build functional, affordable assistive devices. If successful, the technology could bring exoskeletons out of the lab and into everyday life, where they have long been promised but rarely seen.
