Building Smarter Wearable Technologies with AI

Conor Walsh is the Paul A. Maeder Professor of Engineering and Applied Sciences at the Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS). His lab develops wearable technologies that help people move, including robotic devices that physically assist movement, and sensing systems that provide feedback about how someone is moving. We spoke with Walsh about how artificial intelligence can help these technologies work in synchrony with the person, support rehabilitation, and move more rapidly from the laboratory into the real world. The following Q&A was developed from interviews with Walsh. It has been edited for clarity, length, and context.

Q: How is AI changing the way you develop wearable technologies?

A: When we started our research group about 14 years ago, we were largely a hardware-focused group. We were building soft robots that could physically help people move. As that part of the field has advanced, the question has become: how do we leverage all the data these devices can collect, and how do we use that data to make them perform even better?

AI and machine learning give us tools to make sure a device works in synchrony with the person. We can use data from wearable sensors to understand how someone is moving and then optimize the assistance or intervention for that individual.

The objective depends on the application. For a healthy person, it might be helping them walk with less effort. For a stroke survivor, it might be helping them walk faster or more symmetrically. For someone recovering at home, it might be encouraging them to use an impaired arm more often. The common goal is to make the system maximally helpful for that specific person.

Q: Why is personalization so important for wearable robots and assistive technologies?

A: No two people move in exactly the same way, and that becomes especially important in clinical populations. A person’s walking pattern after a stroke, for example, may be very different from someone else’s. Before we can optimize an intervention, we need to understand what that individual actually needs.

We have a collaborative project with Patrick Slade at SEAS, and Lou Awad and Terry Ellis at BU, where we are trying to help stroke survivors walk better. Using the sensors already embedded in a wearable robot, we can estimate things such as walking speed, symmetry between the healthier and less-healthy legs, and how much someone is pushing off the ground. We can then adapt the control parameters to encourage the device to help the person walk in a more natural way.

This is not just about wearable robots. We are also exploring behavioral interventions. In another project with Susan Murphy at SEAS, stroke survivors wear a smartwatch at home that monitors how they use their arm during daily activities. The longer-term goal is to personalize when the system gives feedback, what message it delivers, and how it can best encourage the person to work toward their own recovery goals.

Q: How can AI help you understand movement outside the laboratory?

A: In the lab, we can use motion-capture systems and force-measuring equipment to collect very rich information about how someone moves. But we cannot send people home wearing dozens of sensors.

A major focus of our work is figuring out how to estimate that rich movement data using sparse, minimal sensors in the real world. For a runner, that could mean understanding braking and propulsion forces or aspects of running form. For a stroke survivor, it could mean estimating walking symmetry or how effectively the impaired leg is pushing off. For someone with Parkinson’s disease, it could help us understand how their walking changes during daily life.

To do that well, we also need strong datasets. We are building datasets that combine detailed laboratory measurements with data from wearable sensors for people with stroke and Parkinson’s disease and for people performing activities of daily living with their arms. Those datasets help us connect the high-resolution information we can measure in the lab with the much sparser information we can collect in the real world.

Q: What opportunities does AI create for rehabilitation?

A: Physical and occupational therapists are a critical part of a person’s recovery, but they cannot be with someone all the time. It is generally accepted that more rehabilitation practice and greater intensity can be beneficial, yet there are practical limits on how many sessions a person can receive.

We see AI agents as a potentially useful tool for therapists. A therapist could work directly with a patient and also use an AI-supported system to guide activities between visits. The system could collect data about how the person is doing, deliver reminders or feedback, and help personalize the rehabilitation program over time.

The goal would not be to replace the therapist. It would be to augment the therapist’s ability to support more people and interact with them more frequently. The therapist would still help determine the person’s goals and control how the system is being used.

Q: What are the challenges of using AI in technologies worn on the body?

A: These systems need to be safe. A machine-learning model is never going to be perfectly accurate, and we need to understand what happens when its prediction is wrong.

That is especially important with a wearable robot. If a language model gives you a bad answer, you may be able to ignore it. But if you are wearing a device and it makes an incorrect decision about how to assist you, that could be a problem.

One important area is understanding the certainty of a prediction. If the system recognizes that it has low confidence, it might switch into a more conservative safety mode rather than trying to provide the maximum possible assistance. In that situation, doing no harm is more important than trying to do as much good as possible.

Q: What is the future of AI-powered wearable technology?

A: One opportunity is to make the relationship between people and wearable devices more transparent. These technologies are very intimate — you are wearing them and interacting with them — so people should be able to understand what the device is doing.

In the future, a person might ask a device why it suddenly feels different, and the system could explain that it was observing their walking pattern and adjusting its control parameters. A user or therapist might also be able to describe a goal, such as improving body position while walking, and the system could translate that into an action the device can perform.

AI could also accelerate the development of these technologies. It can help researchers write software, build interfaces, and create higher-fidelity prototypes more quickly. That matters because our goal is not simply to demonstrate something once in the lab. We want to build robust systems that can collect data, operate reliably, and help people in the real world.