How would you lay out a whole-body sEMG suit? We tried and are not convinced
PavelKamenik
czech republic
We are designing an open-hardware sEMG suit, still on paper, with an AI assistant, Claude. The electronics are designed; what we cannot settle is where the electrodes go. Fixed so far: 280 bipolar channels, every one differential (P against N, no reference), on 16 modules of 32 electrodes (the neck's 16), gloves and socks of 16 each, and 100 IMUs on the bones. Rules we keep: short pairs near the heart and none across the sternum, across the spine only diagonally, nothing in the hair, nothing where the body is pressed (palm, sole, seat), electrodes at least 1.5 cm apart. A learned model takes the envelopes, with the IMUs as its teacher; no motor-unit decomposition.
On an anatomical atlas we tried one electrode per muscle belly, two rows of oblique pairs round each limb, and on the trunk pairs from one muscle to its neighbour and pairs scattered by a search to survive the suit shifting a few centimetres. None convinced us; each try came out different, and honestly we do not know how to judge a layout. Nobody puts a suit on the same way twice, so we want one that tolerates a shift.
What we think so far: a pair's lean counts against the fibres, not the body, so on an oblique muscle a level pair is already oblique; a pair across several neighbouring muscles only where it pays, and on the limbs probably single muscles or blocks of muscles that work together; the erectors as six blocks at least.
How would you lay it out? Where single muscles and where blocks, how to divide the erectors, and where on the trunk to tell the rectus, the obliques and the erectors apart? We can share pictures of what we tried.
Comments
One useful layout criterion would be what the EMG adds after refitting. Your SIGNALS.md currently gives the model IMU inputs while using IMU-derived joint angles as targets. I’d compare IMU-only, EMG-only and combined inputs on the same task; otherwise a good joint-angle score could tell you very little about the electrode layout.
For the first-arm comparison, train on earlier fittings and test on an entirely new putting-on of the sleeve. If short recalibration is allowed, reserve those calibration data before scoring the remaining trials, with the same time allowance across layouts. That gives you a practical way to rank layouts for shift tolerance, without pretending an atlas alone settles it.
Thanks, that is a fair catch. With the IMU steps in the input and the IMU-derived joint angles as the target, a good score says nothing about the electrodes. For the first arm we will score every layout on the same task three ways, IMU only, EMG only and both, and rank the layouts by the EMG-only score and by what the EMG adds to the IMUs. Training on earlier fittings and testing on a fresh putting-on of the sleeve, with any recalibration data held out and the same time for every layout, is the protocol we will write down.
One question: would you score the joint angles at all, or rather what the IMUs cannot see, such as a short look-ahead (the muscles fire before the limb moves) or force held without movement?
I'd keep joint angles as the baseline, then add one task that actually needs EMG. A held contraction against a fixed load cell would be a useful first test: measure force independently, vary force at the same posture, and include relaxed trials. Otherwise the model can learn posture or the instruction rather than force.
For look-ahead, fix the horizon before training and allow each input only samples available at prediction time, including the filtering and envelope calculation. Compare EMG plus IMU with an IMU-history predictor at that same horizon; a smooth movement may already be predictable from its recent trajectory. Keep the fresh-fitting split for both tests. I wouldn't use EMG amplitude itself as the force label, or treat the EMG-to-movement lead as a single fixed delay.