Blog

Exoskeletons today look like something straight out of science fiction. But the reality is that they are nowhere near as healthy as their fictional counterparts. They are quite fickle, and it takes many hours to manually develop the software policies that govern how they work—a process that must be repeated for each individual user.

To bring the technology a little closer AvatarSkel Suits, or Warhammer 40k power armor, a team at the University of North Carolina’s Biomechatronics and Intelligent Robotics Lab used AI to build the first universal exoskeleton that supports walking, running and climbing stairs. Most importantly, its software adapts to new users without the need for user-specific settings. “You just wear it and it works,” said Hao Su, associate professor and co-author of the study.

Custom robots

An exoskeleton is a robot you wear to assist your movements – it makes walking, running and other activities less strenuous, in the same way that an e-bike adds extra watts to those you generate yourself, making pedaling easier. “The problem is that exoskeletons have a hard time understanding human intentions, whether you want to run, walk or climb stairs. This is solved with motion recognition: systems that recognize a person’s movement intentions,” says Su.

Building these motion recognition systems currently relies on complex policies that define what the actuators in the exoskeleton should do in every possible scenario. “Let’s go for a walk. The current state of the art is that we put the exoskeleton on you and you walk on a treadmill for an hour. Based on that, we try to adapt his work to your individual range of motion,” explains Su.

Building hand-crafted control policies and performing lengthy human trials for each user makes exoskeletons super expensive, with prices reaching $200,000 or more. So, Su’s team uses AI to automatically generate control policies and eliminate human training. “I think within two or three years, exoskeletons priced between $2,000 and $5,000 will be absolutely feasible,” Su claims.

His team hopes that these savings will come from developing a control policy for the exoskeleton using a digital model rather than living, breathing humans.

Digitizing robo-assisted humans

Su’s team began by building digital models of the human musculoskeletal system and an exoskeleton robot. They then used multiple neural networks that controlled each component. One worked with the digitized model of a human skeleton, driven by simplified muscles. The second neural network controls the model of the exoskeleton. Finally, the third neural network was responsible for imitating movement — basically predicting how a human model would move wearing the exoskeleton and how the two would interact with each other. “We trained all three neural networks simultaneously to minimize muscle activity,” says Su.

One problem the team faced is that exoskeleton research typically uses a performance metric based on a reduction in metabolic rate. “However, humans are incredibly complex, and it is very difficult to build a model with enough fidelity to accurately simulate metabolism,” explains Su. Fortunately, according to the team, the decrease in muscle activations is quite closely related to the decrease in metabolic rate, so it keeps the complexity of the digital model within reasonable limits. Training the entire human-exoskeleton system with the three neural networks took approximately eight hours on a single RTX 3090 GPU. And the results were record breaking.

Bridging the gap between sim and real

After developing the controllers for the digital exoskeleton model that were developed by the neural networks in a simulation, Su’s team simply copied the control policy of a real controller driving a real exoskeleton. They then tested how an exoskeleton trained in this way would perform with 20 different participants. The average reduction in metabolic rate when walking was over 24 percent, over 13 percent when running, and 15.4 percent when climbing stairs—all record numbers, meaning their exoskeleton beat every other exoskeleton ever made in every category.

This was achieved without requiring any adjustments to fit individual gaits. But the magic of neural networks doesn’t end there.

“The problem with traditional, hand-crafted policies was that it just said ‘if walking is detected, do one thing; if you are found to be going faster, do another thing.’ These were [a mix of] state machines and switching controllers. We introduced continuous end-to-end control,” says Su. What this continuous control meant was that the exoskeleton could follow the human body as it made smooth transitions between different activities—from walking to running, from running to climbing stairs, and so on. There was no abrupt mode switching.

“In terms of software, I think everyone will soon be using this neural network-based approach,” says Su. To improve exoskeletons in the future, his team wants to make them quieter, lighter and more comfortable.

But the plan is also for them to work for people who need them the most. “The limitation now is that we tested these exoskeletons with healthy participants, not people with gait impairments. So what we want to do is something that they did in another exoskeleton study at Stanford University. We will capture a one-minute video of you walking and build a model based on that to individualize our overall model. This should work well for people with disabilities such as knee arthritis,” claims Su.

Nature, 2024. DOI: 10.1038/s41586-024-07382-4