Project overview
AlphaHand is a Muse 2 EEG research project testing whether consumer-grade neural signals can produce auditable evidence for action-state decoding, active-finger identification, and robotic-hand command gating.
Featured run
2-m16 · March 19, 2026
March 19 checkpoint with the current per-finger actuation defaults.
Primary holdout
84.66% joint accuracy
Holdout diagnostics keep joint accuracy, rest behavior, and applicability separate.
Public corpus
2 runs · 4,953 windows
Published bundles include metrics, figures, reports, and reproducibility artifacts.
Replay evidence
95.37% would-send precision
Pseudo-live replay keeps deployment behavior separate from offline split accuracy.
Each run turns Muse 2 EEG into short temporal windows, then evaluates whether a model can separate the REST/OPEN/CLOSE action-state label and identify the active finger when movement is present.
The deployment path adds a finger-applicability head, confidence thresholds, smoothing, stability checks, and same-finger hold logic before any prediction is treated as a robot-hand command.
The site separates saved test-split metrics, primary holdout diagnostics, event-level summaries, and pseudo-live replay because each answers a different question about model behavior.
April 3 remains the stronger historical offline checkpoint, while the current public bundle uses March 19 because its deployment replay is more reliable under precision, event-hit, latency, and false-actuation checks.
Start here
Reviewers can start with the current results, engineers can inspect the pipeline, and researchers can follow the selection audit back through the dated update trail.
Results
Headline metrics, figures, per-run bundles, and evidence labels.
How it works
Capture, preprocessing, model heads, replay, and actuation gates.
Deep dive
Model-selection audit, tuning history, and March 19 vs April 3 tradeoffs.
Visualizations
Interactive latent-space views for inspecting learned structure.
Poster and paper status
Public ISEF poster board PDF and private manuscript status.
Updates
Dated research notes documenting how the project evolved.
The poster board displayed at ISEF is public on this page as paper.pdf. The full manuscript remains in progress and private as of Sunday, May 17, 2026.
Open poster and paper statusThe strongest extensions are practical and testable: reproduce a published bundle, improve same-session validation, add subject coverage, evaluate safer control gates, or contribute clearer reporting around model-selection tradeoffs.