HactionComplet Online BCI Control

This project now has a clear online control pipeline:

1. Read EEG from LSL
2. Build sliding/overlapping windows
3. Run model prediction
4. Send UDP commands to the game

The main script for this is online_windowing.py.

Project Layout

• online_windowing.py: online LSL windowing + model inference + optional UDP send
• lsl_connect.py: full-featured EEG bridge with recording, diagnostics, and game control
• model_Data28_BIS.pkl: CSP+LDA model used for prediction
• test_lda.py: smoke test for model loading/prediction
• fake_lsl_eeg.py: fake EEG LSL stream generator for testing
• Hacktion_game-main/: game with BCI integration (UDP receiver on 127.0.0.1:5005)
• RECORDING_FIX.md: troubleshooting guide for EDF recording issues

Install

pip install -r requirements.txt

Quick End-to-End Run (Real Game Control)

Terminal 1 (start game):

python Hacktion_game-main/Main.py

Optional model selection:

python Hacktion_game-main/Main.py --select-model

Or force one model:

python Hacktion_game-main/Main.py --model-path Hacktion_game-main/Model_simple/model_0010.pkl

Terminal 2 (start online BCI bridge):

python online_windowing.py \
  --stream-type EEG \
  --window-seconds 1.0 \
  --overlap-seconds 0.8 \
  --fixed-fs-hz 256 \
  --model-path model_Data28_BIS.pkl \
  --udp-host 127.0.0.1 \
  --udp-port 5005

No EEG Hardware? Use Fake LSL Stream

Terminal A:

python fake_lsl_eeg.py --stream-name FakeEEG --stream-type EEG --sample-rate 250 --channels 8

Terminal B:

python online_windowing.py \
  --stream-name FakeEEG \
  --window-seconds 1.0 \
  --overlap-seconds 0.8 \
  --fixed-fs-hz 256 \
  --car \
  --bandpass-low-hz 1 \
  --bandpass-high-hz 40 \
  --model-path model_Data28_BIS.pkl \
  --udp-host 127.0.0.1 \
  --udp-port 5005

Command Mapping to Game

Default label-to-UDP payload map is:

• left_hand -> "-1"
• right_hand -> "1"
• idle -> "0"

You can override with:

--udp-label-map '{"left_hand":"-1","right_hand":"1","idle":"0"}'

or pass a JSON file path.

Useful Flags

• --window-seconds: window duration
• --overlap-seconds or --stride-seconds: sliding control
• --fixed-fs-hz: force model input sample rate (useful if model expects fixed rate)
• --car: apply common average reference
• --bandpass-low-hz, --bandpass-high-hz: optional FFT bandpass
• --min-confidence: minimum confidence required before UDP send
• --health-interval-seconds: periodic LSL health logs (samples/sec, windows/sec, stale/live)
• --no-samples-warn-seconds: warning threshold if stream goes silent

Live Validation In Game

When you run python Hacktion_game-main/Main.py, the game now displays:

• BCI UDP: WAITING/LIVE/STALE
• total UDP packets and valid commands received
• last decoded BCI command with age
• keyboard fallback state

Press K in game to toggle keyboard fallback ON/OFF.
Set it to OFF to verify headset-only control.

Calibration -> EDF -> New Weights

From the game menu, run calibration. The calibration flow now:

1. Records EEG from LSL and saves calibrage_recording_YYYYMMDD_HHMMSS.edf at project root.
2. Tries to train a new CSP+LDA model from this EDF.
3. Saves model weights under Hacktion_game-main/Model_simple/model_<edf_name>.pkl.
4. Automatically sets this new model as active for the next game run in the same session.

Useful env options:

• BCI_CALIBRATION_TRAIN=0 to skip auto-training after calibration
• BCI_STREAM_NAME=... / BCI_SOURCE_ID=... to target a specific EEG stream during calibration and game

Validation Commands

Model smoke test:

python test_lda.py --model-path model_Data28_BIS.pkl --channels 8 --samples 256 --batch 3

List available LSL streams:

python online_windowing.py --list-streams