MIT mechanical engineers have built a rehabilitation robot for stroke patients that learns to assist by observing physical therapists at work. Instead of programming the robot with a fixed motion, the team recorded therapists guiding patients through exercises — including how they respond to touch, force, and resistance — and trained a generative model on that interaction data. The robot then adjusts the level of assistance based on the patient’s own effort.
The team’s framing is worth quoting. The goal, says lead researcher Johannes Lachner, is not to replace therapists but to extend their reach. In an ongoing clinical study in Munich, therapists wearing force-sensing gloves are being recorded treating patients, and the next phase aims to build therapist-specific models — a robot that carries one clinician’s particular style of care, tested with patients who previously received manual therapy from that same clinician.
Strip away the robotics, and this is the same idea that runs through everything we do: identify the expert, capture what they actually do rather than what the textbook says, and build a system that carries that expertise forward. The therapist trains the robot. The transplant surgeon defines the rules. The senior engineer audits the model. In each case, the human is the source of the intelligence, and the technology amplifies it. Ford learned this year what happens when you skip that step.
MIT News, “Personalized physical therapy: Stroke rehabilitation powered by AI” (Aug. 5, 2026):
https://news.mit.edu/2026/personalized-physical-therapy-stroke-rehabilitation-powered-by-ai-0805
Paper in IEEE Transactions on Robotics, “Diffusion-Based Impedance Learning for Contact-Rich Manipulation Tasks”: