Seán Barrett

Research

I work on reading muscle signals well enough to drive a hand. This preprint, with William Hartley, is the first of it in writing.

Preprint

A Leak-Safe Within-User Benchmark for Compact Surface-EMG Grasp Decoding, with a Causal Sequence-Reasoning Decoder

Authors
Seán Barrett (lead author), William Hartley
Venue
Preprints.org, preprint, not yet peer reviewed
Posted
Licence
CC BY 4.0

In plain English

Plenty of research on reading muscle signals for prosthetic hands reports impressive accuracy, measured offline. Look closer and the models are often too big for the small chips inside a real hand. The testing often lets a model see data it is later scored on. And a single headline score can hide how the decoder behaves moment to moment. This paper is mostly about testing honestly. On a public dataset, we calibrate on each person's earlier attempts at a grip and test on a later one the model has never seen. We report accuracy moment by moment and per whole attempt, and how often it fires when the hand should stay still. Then we test a decoder small enough for a microcontroller, which reasons about how grips follow one another. The code is open.

Read the preprintThe code

The idea, in one figure

Hand: open

A muscle that works harder makes a bigger, noisier electrical signal. Flip the negative half up (rectify it), smooth it into an envelope, and compare that with a threshold. Above it, the hand closes. That is the basic idea behind how a myoelectric hand is controlled.

Simulated signal. The general idea only, not how any particular hand does it.

Cite it

@misc{barrett2026leaksafe,
  title     = {A Leak-Safe Within-User Benchmark for Compact Surface-EMG Grasp Decoding, with a Causal Sequence-Reasoning Decoder},
  author    = {Barrett, Se{\'a}n and Hartley, William},
  year      = {2026},
  month     = jun,
  publisher = {Preprints},
  doi       = {10.20944/preprints202606.1365.v1},
  url       = {https://doi.org/10.20944/preprints202606.1365.v1},
  note      = {Preprint}
}