MIT-Linked Team Develops Hybrid Human-AI Rehab System to Help Stroke Survivors Amid Therapist Shortages

Each year, strokes leave an estimated 5 million people living with lasting disabilities, often requiring months of carefully guided rehabilitation to regain basic movements and independence. But as demand rises, access to care is increasingly constrained by a growing shortage of physical therapists—making it harder for many patients to receive enough supervised therapy to recover as fully as possible.

A research team with ties to MIT is pursuing a hybrid solution that pairs the judgment of human clinicians with the consistency and scalability of robotics. Johannes Lachner, formerly a postdoctoral researcher at MIT and now an assistant professor at Purdue University, and Noah Geiger, a former visiting student at MIT, have developed a robotic rehabilitation system designed to learn directly from the way therapists work with patients.

The device relies on two technical pillars: generative artificial intelligence and real-time force feedback. In practice, the robot can sense how much resistance or assistance a patient is providing during an exercise—then adjust its support on the fly. The generative AI component is trained using demonstrations from professional therapists, enabling the system to infer how a clinician might respond to different movement patterns, difficulties, or levels of fatigue.

The aim is to deliver “just the right amount” of help during rehabilitation tasks. These can include foundational movements such as reaching for an object, lifting an arm, and practicing everyday functional activities that often become challenging after a stroke. By modulating assistance in real time, the system is meant to encourage patients to do as much work as they can safely manage, while still preventing frustration, unsafe strain, or poor movement habits.

Lachner stresses that the technology is intended as support for clinical practice, not a substitute for trained professionals. “Not to replace therapists, but to extend their reach,” he says—framing the robot as a way to scale expert care to more patients and potentially increase the intensity or duration of therapy when staffing is limited.

Early tests using the robotic platform indicate the core concept is workable. Lachner reports that initial experiments with the robot have already shown the approach is feasible, providing a foundation for more ambitious evaluations.

The next stage, he says, is to build therapist-specific models—systems tuned to individual clinicians’ styles and techniques—rather than relying on a one-size-fits-all approach. The team also plans a longer-term clinical study that would follow the same patients who previously underwent traditional, hands-on therapy, allowing researchers to compare outcomes over time and assess how well the learned strategies translate to real rehabilitation settings.

If the system proves reliable in clinical use, Lachner sees potential beyond medicine. Because the robot is designed for “contact-rich” interactions—situations where touch, force, and fine motor control matter—similar methods could be adapted for industrial environments or collaborative workplaces where humans and robots must safely coordinate physical tasks.

Why it matters

With millions facing long-term disability after stroke and too few therapists to meet demand, tools that safely increase therapy access could have wide impact for patients and health systems alike. The key questions now are whether AI-guided robots can match the adaptability of skilled clinicians and whether personalized “therapist-specific” models improve outcomes. Results from longer clinical studies will be critical in determining whether this approach can move from feasibility to routine care.