Myo-Amp

Three sEMG pads read the muscles in a forearm, a model decides which grip you meant, and cable-driven linear actuators close your fingers around the object. Five of us built it over a semester. I owned the firmware, which is where all of it had to meet.

for this project, the firmware was the integration

Timur Gray, mechanical · Brandon Spiller, electrical · Anna Du, machine learning · Masha Arakcheeva, team management

• Three sEMG channels sampled round robin on one ADC, so no muscle waits its turn
• A shared data object living on both cores of the RP2040, handed across as a pointer through the hardware FIFO
• PWM and direction for three linear actuators, with the drivers idled between moves
• TensorFlow Lite Micro running on the second core, with Anna's trained model compiled into the binary

One core reads the sensors, the other drives the motors. Put both on one core and the signal waits behind the motion, and a grip that lags is a grip nobody trusts.

They share one object. Core zero builds it and passes the pointer across the hardware FIFO, core one catches it and casts it back. From then on both cores are writing into the same place, which is the part you have to be careful about.

16 MHz on the Arduino we started on. 133 MHz and two cores on the Pico.

The Pico's own voltage reference was too noisy to measure muscle activation against, and it could not push enough current for three sensors anyway. So the sensors got their own clean three volt line, and that same line went into the ADC reference pin.

A reading is only as good as the thing you compare it against.

The model picks between three grips and two kinds of not moving: open hand, air pinch, full fist, resting, and still resting. Anna Du trained it. My job was getting it onto the chip and turning its answer into motors.

The model path is written and it is not connected. The firmware samples, the classifier has somewhere to put its answer, and the second core runs a fixed sweep instead of the grip it picked.

// instance->get_linear_actuator(0).drive(instance->get_model_value(0));
// insert ml read code

still sitting in main.cpp

The device underperformed. One motor driver was carrying two linear actuators, went past what it could supply, and broke. The replacement never shipped, so the build finished on a mismatched pair and reliability went with it.

The current budget was the mistake, and we found it too late to re-spec.