ThoughtLink: Brain-to-Robot Control

Hack Nation 4th Global AI Hackathon | Feb 7-8, 2026

Real-time motor intent decoding from EEG+fNIRS brain signals.

---

Results

• 94.6% accuracy (5-class motor imagery)
• 3.5ms average latency (14x under 50ms limit)
• All classes >92% balanced performance

---

Quick Start

pip install -r requirements.txt
python data/download.py
python train.py
python predict.py data/cache/robot_control/data/<sample>.npz

---

Approach

Competitive Edges:

• M2 variance moment extraction (highest cortical selectivity)
• Short-channel regression (removes scalp noise)
• Sliding window augmentation (10x training data)

Architecture:

• Simple RBF-SVM (fast, scalable)
• Multimodal fusion: EEG bandpower + fNIRS moments
• PCA dimension reduction + confidence thresholding

---

Dataset

5 classes: Right Fist, Left Fist, Both Fists, Tongue Tapping, Relax

• EEG: 6 channels @ 500Hz (frontal placement: FCz, CPz critical)
• TD-NIRS: 40 modules @ 4.76Hz (full head coverage)

---

Integration

The predict() function outputs:

{
    'command': 'RIGHT_FIST',
    'confidence': 0.98,
    'latency_ms': 3.4
}

Ready to plug into robot simulation using a connector script.

---

Why It Works

• The 6 EEG channels miss motor cortex (C3/C4), so treating the 40-module fNIRS array as primary signal source.
• Short-channel regression + M2 variance extracts cortical signals.
• Sliding windows give the simple SVM enough data to learn robust patterns without overfitting.