Emerging Interfaces

A Rehab Robot That Reads Your Muscles, Not Just Your Joint Angles

Most robot-assisted rehabilitation systems prescribe where a limb should move. A new transformer-based model instead targets which muscles should be firing — reaching the intended activation pattern with just 15% error in live testing.

Robot-mediated physical rehabilitation has a quietly persistent problem: most systems are built to guide a limb through a specific path — a kinematic trajectory — rather than to target the actual muscle activation that therapy is supposed to be training. A new model, accepted for presentation at the 2026 Conference on Robot Learning (CoRL), flips that target. Instead of prescribing where a joint should go, it prescribes the contact-force trajectory that should produce a specific, personalized pattern of muscle activation — measured directly via surface electromyography (sEMG).

Reading muscle, not just motion

The system uses a transformer encoder-decoder with rotary positional embeddings to build a latent representation of what the researchers call “trajectory-relevant and user-specific neuromotor factors” — in plainer terms, a model of how a specific person’s muscles respond to a specific therapeutic task. It then augments its trajectory predictions with flow matching, producing a distribution of plausible force prescriptions biased toward realistic, physically sensible outputs rather than a single rigid answer. Trained and evaluated on 16 healthy participants across 15 isometric trajectory-tracking tasks, the model generalizes to both held-out tasks and held-out users with under 8% mean trajectory prediction error, running fast enough for effectively instantaneous inference during an actual session.

It worked when it mattered: live testing

The more interesting result sits outside the benchmark numbers: an initial live test running the model during actual rehabilitation tasks found participants reaching their target neuromotor activation pattern with only about 15% error accumulated across the whole task — evidence that a model trained on clean lab data can hold up once it’s actually driving a robot’s force output in real time, rather than just scoring well on a held-out test set. The team has open-sourced the data and code as part of the OpenRobotRehab project.

Why targeting muscles instead of motion matters

Robot-assisted rehabilitation devices have had a rocky track record on outcomes specifically because kinematic-trajectory systems can guide a limb through the “correct” path without actually engaging the muscles recovery depends on — a patient’s arm can trace the right arc while barely working the muscle group therapy is meant to target. A system that personalizes its force output around a measured, individual neuromotor signal rather than a generic motion template is a meaningfully different kind of human-machine interface: one where the “input” the machine is responding to is the body’s own electrical activity, not just its position in space. That’s the same broader shift — reading signals the body already produces and feeding them directly into a responsive system — that runs through most of the emerging-interface work this site covers, just applied here to clinical recovery rather than creative control.