MIT Rehabilitation Robot Learns From Physical Therapists to Support Stroke Patients

Mechanical engineers at MIT have built a physical therapy robot that, instead of replaying a preprogrammed motion, learns from physical therapists how to respond to touch, muscle resistance and a patient's real-time effort. The system is meant to help stroke patients at a time when the number of people needing rehabilitation is growing faster than the number of available therapists.

The robot has two arms and works in direct physical contact with the patient. Rather than moving along a rigidly programmed trajectory, the system adjusts its force and the way it provides support based on how much effort the patient is able to put into the exercise at any given moment. That sets it apart from earlier generations of rehabilitation robots, which mostly offered repetitive movement sequences identical for every patient.

Learning touch, not just motion

At the heart of the system are diffusion models, the technology best known for generating images in tools like ChatGPT, but here applied not to pixels, instead to the robot's physical behavior. The model learns how firmly and in what way to support a patient's arm in response to touch, resistance and level of effort, rather than simply replaying a memorized movement trajectory.

While most generative AI models in robotics focus on motion, ours is one of the first to learn physical interaction, meaning how to respond to touch, force and resistance - Noah Geiger, MIT

The team trained the model on data from teleoperation experiments in which healthy participants performed rehabilitation movements, such as lifting and reaching with an arm, at varying levels of effort. The researchers describe this dataset as a kind of movement obstacle course, covering a contact-rich repertoire of interactions that can occur between a therapist and a patient.

Therapists train their own robot

The key difference from earlier approaches is that physical therapists themselves can train the robot they work with. In an ongoing clinical study in Munich, therapists wear gloves equipped with force sensors during real sessions with patients. The data collected this way feeds into an AI model that learns the individual therapist's working style and translates it into the robot's behavior.

The core innovation of our system is that physical therapists can train their own robot using AI, providing personalized, scalable support tailored to each patient's needs - Johannes Lachner, MIT

This approach is meant to address a problem facing healthcare systems worldwide: a growing number of stroke patients alongside a shortage of qualified physical therapists. The robot isn't meant to replace therapists, but to let them handle more patients by shifting some of the repetitive yet precision-demanding work onto a machine trained to mimic each therapist's individual approach to therapy.

Uses beyond stroke rehabilitation

MIT researchers say the same technology could find applications beyond stroke rehabilitation, including post-surgical recovery, maintaining muscle strength in older adults, and collaborative robot workstations in industry, where a machine must sense resistance and adjust its applied force to the task and the operator.

For Poland, where access to post-stroke physical therapy outside major centers can be limited, this kind of technology remains for now at the clinical-trial stage abroad, and there are no announcements yet about deployment in Polish rehabilitation facilities. Still, the project points to a direction AI-assisted rehabilitation may take in the coming years: not a therapist replaced by a ready-made exercise program, but a machine that learns the specific working style of a specific specialist.

The team's paper, titled Diffusion-Based Impedance Learning for Contact-Rich Manipulation Tasks, was published in the journal IEEE Transactions on Robotics in 2026 and details the model's technical design along with results from impedance-learning experiments in contact-rich manipulation tasks.