Project B: Control for a real user
EMG or EEG control of a wheelchair, a robot arm, or a drone, for a specific person. Drone-by-brain is a demo. Wheelchair-by-muscle for someone who cannot use a joystick is a life.
You are skimming: the title, the first figure, and the short version. Switch to Read in the header for the full page, or Deep to open every deep dive.
The eight-channel armband recognized your gestures. The motor imagery classifier moved a bar, sometimes. Now the question is who could use either. A person with a high spinal cord injury who cannot grip a joystick may have shoulder or neck EMG. A person with cerebral palsy may have reliable EMG at one site and unreliable movement everywhere else. Find the person, find the signal they have, and build the controller for it. The result will be less flashy than a drone and will change how someone gets across a room.
A person with a C4 spinal cord injury cannot move their arms or hands. Which control signal is most likely available and reliable?
EMG from muscles above the injury: the sternocleidomastoid and trapezius in the neck and shoulder, or facial muscles, plus head movement. These are large, reliable, learnable signals, which is why commercial wheelchair controls for high injuries use head arrays and chin joysticks. Motor imagery EEG is the last resort, not the first, because its reliability is a fraction of EMG’s. Building the reliable thing first is what the person needs; the EEG version is research.
Find the person and the signal
Through the rehabilitation hospital’s occupational therapists, an independent living centre, or the Utah Assistive Technology Program. The occupational therapist knows what controls have been tried and why they failed. Map the person’s reliable voluntary signals: which muscles they can contract on command, how repeatably, for how long before fatigue. Record each candidate site with the armband hardware. The signal with the highest repeatable signal-to-noise and the least fatigue wins, whatever it is.
The controller
Two to four discrete commands is the realistic starting point: forward, stop, left, right; or open, close, and a mode switch. Map them to gestures the person can make reliably: a short contraction, a long one, two in a row. Add a confirmation for anything dangerous. Give continuous visual and auditory feedback of what the system thinks it saw, so the person can correct. Latency under 300 ms.
For a wheelchair, do not drive the motors yourself. Commercial chairs accept alternative input devices through a standard interface; your controller emulates a switch array or a joystick voltage through that interface, inheriting the chair’s own safety systems. For a robot arm, the same: command a well-tested arm’s API, do not build the arm.
Iterate with the person
Session one: the person uses it for ten minutes in a safe space while you watch. What confused them, what fatigued them, what did the system get wrong. Session two: the fixed version. Repeat weekly. Measure the things that matter to them: time to cross the room, number of unintended stops, how long they can use it before tiring. Then measure the things the FMEA cares about: false activations per hour.
The EEG version
If the person has no reliable EMG, or if they and you want to try, add motor imagery or SSVEP as a command channel. Report its accuracy next to the EMG’s, honestly. Most of the time EMG wins by a wide margin. Knowing that, from your own measurements on a real person, is worth more than a drone video.
- Find the person through an occupational therapist. Get onto the protocol. Listen, twice.
- Map their reliable signals with the armband hardware. Pick one or two sites.
- Build a two-command controller with feedback and a kill switch. Test it on yourself, then on a simulator on screen.
- Write the FMEA. Have the therapist read it.
- First live session in a safe space with the chair or arm limited. Iterate weekly.
- Measure task time, unintended activations, and fatigue over four sessions. Report.
- Hand over with a setup guide the caregiver can follow, and a support plan.