• The reported evidence is feasibility: MIT News describes an initial prototype tested with healthy participants, and no stroke-patient results are reported.
• The technical contribution is learning physical interaction (stiffness and damping), not only motion, per the authors' preprint.
• A clinical study with therapists wearing force-sensing gloves is under way in Munich, so the patient-facing claims are untested.
• Neither MIT outlet reports patient counts, outcome measures, or a comparison against manual therapy.
• The stated aim is to let therapists train their own robot, not to replace them.

A stroke-therapy robot that learns from physical therapists targets a real problem, but the public record supports only a feasibility claim. MIT Technology Review reports that Johannes Lachner and Noah Geiger built a system combining generative AI with force feedback. MIT News describes an initial prototype tested with healthy participants and no stroke-patient results. The underlying method is documented in a preprint, which this assessment relies on for technical detail.

What the system does

The system is a dual-arm robot for physical and occupational therapy that adjusts how much help it gives. MIT News says it combines transformer-based diffusion models with real-time force feedback and tailors assistance to each patient's capabilities. Technology Review lists reaching, arm lifting and daily tasks as example exercises. The design goal in both accounts is graded assistance learned from therapists, rather than fixed, pre-programmed motion that ignores how much effort the patient contributes.

The design choice that matters: learning interaction

The technically distinctive claim is that the model learns physical interaction, not only motion. MIT News quotes Geiger saying most generative AI models in robotics focus on motion, while theirs is among the first to learn how to respond to touch. The arXiv preprint describes a diffusion model conditioned on external forces, with an energy-based estimator setting stiffness and damping. That fits therapy, where support firmness matters as much as path, but it is a design rationale, not evidence of benefit.

The evidence is feasibility, not outcomes

Neither article reports efficacy data. Technology Review quotes Lachner saying initial robotic experiments demonstrated feasibility. MIT News describes a prototype tested with healthy participants who performed rehabilitation movements while deliberately varying effort, plus operator-guided contact tasks. It reports no stroke-patient counts, outcome measures or comparison with conventional manual therapy. Without that baseline, nobody can say whether the robot helps, matches or underperforms a human therapist. We did not verify the preprint's publication status.

The clinical study is where the claim gets tested

MIT News reports a clinical study with Professor Cristina Piazza's lab at TUM and the Pfennigparade outpatient centre in Munich. Therapists wear force-sensing gloves and are filmed while treating patients, and the data trains models of each therapist's interaction style. Technology Review says the next phase will develop therapist-specific models and evaluate them in a long-term clinical study with patients who previously received manual therapy. No results are reported, so patient benefit remains open.

Why the staffing argument does not settle it

Technology Review states that strokes leave 5 million people with long-term impairments each year, and that a growing shortage of physical therapists limits their care. That is a reason to want a working tool, not evidence that this one works. Training on individual therapists also creates a dependency: the model can only be as good as the therapist it imitates, and neither article says how that quality would be measured or how failures would be caught.

Verdict: a credible prototype, no basis to adopt

This is a technically coherent research programme with no demonstrated patient benefit. Its likely strength is imitating graded physical assistance; whether that improves recovery, and which patients should not use it, is unreported. Hospital buyers and policymakers should wait for patient outcomes against a manual-therapy baseline in a published study, plus any regulatory filing. Until then it is a result to track, not a system to procure.