ThoughtLink: Brain-to-Robot Fleet Orchestration

Hack Nation 2026 | Challenge 9 | Kernel x Dimensional

BCI systems control one robot. ThoughtLink controls 100.

We decode EEG brain signals into robot movement commands and route them through a three-layer fleet orchestration system. One operator, 100 autonomous Unitree G1 humanoids, intervening only when robots get stuck.

  Raw EEG (.npz)
       |
       v
  +-----------+     +-----------+     +-----------+
  | Bandpass  |---->| Active    |---->| Normalize |
  | 8-30 Hz   |     | Segment   |     | Per Chan  |
  +-----------+     +-----------+     +-----------+
       |
       v
  +-----------+     +-----------+     +-----------+
  | Window    |---->| Feature   |---->| 69 feats  |
  | 1s/0.5s   |     | Extract   |     | PSD+Stat  |
  +-----------+     +-----------+     +-----------+
       |
       v
  +===========+     +===========+
  | Stage 1   |---->| Stage 2   |    (only if active)
  | Rest vs   |     | Direction |
  | Active    |     | 4-class   |
  | RF 79.4%  |     | SVM 27.9% |
  +===========+     +===========+
       |                 |
       v                 v
  +-----------+     +-----------+     +-----------+
  | Confidence|---->| Majority  |---->| Hysteresis|
  | Gate      |     | Vote (5)  |     | Filter(3) |
  +-----------+     +-----------+     +-----------+
       |
       v
  +===========+
  | Fleet     |  100 robots, 10 groups
  | Router    |  Context-aware dispatch
  +===========+

Quick Start

git clone https://github.com/Lekha-Reddy-git-hub/thoughtlink.git
cd thoughtlink
pip install uv && uv venv --python 3.12 && uv sync

# downloads data, trains models, clones BRI (~10 min)
uv run python setup_project.py

# run the fleet demo
uv run python demo/full_demo.py

manual setup

uv run python src/download.py
uv run python src/preprocess.py
uv run python src/features.py
uv run python src/stage1_binary.py
uv run python src/stage2_direction.py
git clone https://github.com/Nabla7/brain-robot-interface.git

The Demo

demo/full_demo.py opens two windows:

Left: MuJoCo G1 humanoid close-up (follows whichever robot is being overridden).

Right: Fleet operations panel. 100 robots in 10 groups. Autonomous robots move in group colors. Stuck robots flash red with diagnosed failure reasons.

Three layers of control:

Layer 1 (individual): Click a brain signal button. Pipeline decodes one EEG recording, sends the action to one stuck robot. Direct but doesn't scale.

Layer 2 (group): Type group 3 left. Sends LEFT to every stuck robot in Group 3. One command, multiple robots.

Layer 3 (context-aware): Type group 3 fix. System diagnoses each robot's failure and prescribes the right corrective action per robot:

• R22 has obstacle on left -> RIGHT
• R25 lost its target -> FORWARD
• R28 failed its task -> BACKWARD

One command, three robots, three different actions. The operator says WHERE. The system decides HOW.

fix all resolves every stuck robot fleet-wide. Press A for auto-play, Q to quit.

Results

Metric	Value
Stage 1 (Rest vs Active)	79.4% cross-subject
Stage 2 (4-class direction)	27.9% cross-subject (chance = 25%)
Inference latency	24ms avg headless
Flicker reduction	92.9%
Fleet size	100 robots, 10 groups
Throughput	~40 robots/sec per core
TD-NIRS fusion	no benefit (zero-variance features)

All metrics are cross-subject (train on 5, test on 1 held out). No within-subject shortcuts.

Why These Design Choices

Two-stage classifier instead of flat 5-class: False triggers (sending FORWARD during rest) are worse than missed commands. The binary gate (79.4%) prevents the direction classifier from running on rest data.

4-class direction at 27.9%: Modest but honest. The challenge asks for all four directions. We kept 4-class rather than dropping to 3-class. Cross-subject generalization with 6 frontal channels is inherently hard since motor cortex channels (C3/C4) are missing.

Smoothing over raw accuracy: For robot control, command stability matters more than per-window accuracy. Majority voting (5 windows) + confidence gating + hysteresis (3 consecutive) = 92.9% flicker reduction.

RF over neural nets: Random Forest with hand-crafted PSD features beats MLP on raw EEG for direction (27.9% vs 24.3%). Domain features capture motor imagery better cross-subject.

TD-NIRS didn't help: All 1,728 NIRS features had zero variance. Hemodynamic response is too slow for motor imagery windows.

Fleet Architecture

Scaling from 1 to 100+ robots:

Layer	Command	Effect
Layer 1	click button	1 action -> 1 robot
Layer 2	group 3 left	1 action -> all stuck in G3
Layer 3	group 3 fix	1 trigger -> N individualized actions
Layer 3	fix all	1 trigger -> all stuck robots fleet-wide

Context-aware dispatch maps failure reason to corrective action: obstacle_left -> RIGHT, lost_target -> FORWARD, failed_task -> BACKWARD, unknown -> STOP. The BCI decoder runs once per decision cycle regardless of fleet size. Group routing is O(N) on pre-decoded intent.

At 26ms per robot, an 8-core machine handles 300+ robots at 1 Hz. Pure CPU, no GPU.

Dataset

KernelCo/robot_control on HuggingFace. 900 recordings, 15s each at 500 Hz. 6 EEG channels (AFF6, AFp2, AFp1, AFF5, FCz, CPz). 5 labels. 6 subjects, 20 sessions.

Modules

Module	What it does
src/download.py	dataset download + validation
src/preprocess.py	bandpass filter, segment, window
src/features.py	PSD + statistical features (69 per window)
src/stage1_binary.py	rest vs active classifier (RF)
src/stage2_direction.py	direction classifier (SVM, 4-class)
src/smoothing.py	majority vote + confidence gate + hysteresis
src/pipeline.py	end-to-end ThoughtLinkPipeline
src/integration.py	BRI robot integration (MuJoCo + fallback)
src/cnn_baseline.py	MLP vs RF comparison
src/explore_nirs.py	TD-NIRS multimodal fusion test
src/temporal_embedding.py	intent evolution PCA trajectories
demo/full_demo.py	100-robot fleet demo
demo/run_demo.py	single file decode + robot
demo/scalability_demo.py	fleet scaling benchmark

Limitations

• 6 channels vs 64+ in research BCI systems
• 4 of 6 channels are frontal, not over motor cortex (C3/C4 missing)
• cross-subject generalization without calibration is fundamentally hard
• 4-class direction at 27.9% is above chance but not reliable for fine control
• offline processing (pre-recorded .npz, not live EEG stream)

See ARCHITECTURE.md for full design decisions.

Built With

Python 3.12, scikit-learn, scipy, MuJoCo, Unitree G1, Kernel Flow EEG, brain-robot-interface (Nabla7)

Acknowledgments

Built at Hack Nation 2026. Dataset from KernelCo. Simulation via Nabla7/brain-robot-interface.