Hybrid VTOL — Neuro-Adaptive Sensor Fusion

Research-grade implementation of a Neuro-Adaptive EKF for a Hybrid VTOL aircraft.
Companion Computer (Raspberry Pi 5) ↔ Flight Controller (Cube Orange+).

---

Analysis and Results

Active FDNN

Inactive FDNN

Architecture

Cube Orange+ (PX4 EKF2)              Raspberry Pi 5
────────────────────────              ──────────────
  EKF innovations ── MAVLink ──▶  InnovationMonitor
                                       │  deque(W=64), decimate every 8th
                                  /neuro/features (24-dim)
                                       │
                                  DNNInference
                                       │  neuro_adapter.tflite (INT8, 10 KB)
                                       │  Eq 30: R_adapt = R₀ + α · ΔR
                                  /neuro/covariance_correction
                                       │
                                  CovarianceInjector
                                       │
  EKF2 R matrix ◀── MAVLink ──────────┘

Prerequisites

Tool	Version
Python	≥ 3.10
uv	latest
ROS 2	Humble
TensorFlow	≥ 2.16

---

1 — Install Dependencies

cd hybrid_vtol_neuro_fusion

# Create virtual environment
uv venv --python 3.10
source .venv/bin/activate

# Install project + dev tools
uv pip install -e ".[dev,viz]"

# Install TensorFlow (training, host only)
uv pip install tensorflow

2 — Generate Synthetic Training Data

python -m ml_core.generate_data \
    --n-flights 50 \
    --duration 120 \
    --seed 42

Output: ml_core/data/features.npy and ml_core/data/labels.npy.

3 — Train the Supervisor DNN

python -m ml_core.train_model \
    --data-dir ml_core/data \
    --epochs 100 \
    --batch-size 128

Output: ml_core/models/supervisor_dnn/ (SavedModel) and .keras.

4 — Quantize to INT8 TFLite

python -m ml_core.quantize \
    --model ml_core/models/supervisor_dnn \
    --data-dir ml_core/data \
    --output ml_core/models/neuro_adapter.tflite

Output: ml_core/models/neuro_adapter.tflite (~10 KB, full INT8).

5 — Launch the ROS 2 Stack

# Source ROS 2
source /opt/ros/humble/setup.bash

# Build the workspace
cd src/neuro_adaptive_fusion
colcon build --packages-select neuro_adaptive_fusion
source install/setup.bash

# Launch all 3 nodes
ros2 launch neuro_adaptive_fusion system.launch.py \
    model_path:=ml_core/models/neuro_adapter.tflite \
    alpha:=1.0

Project Structure

hybrid_vtol_neuro_fusion/
├── pyproject.toml
├── ml_core/
│   ├── generate_data.py          # Synthetic innovation sequences
│   ├── train_model.py            # Supervisor DNN training
│   ├── quantize.py               # INT8 TFLite quantization
│   └── models/                   # Saved artifacts
├── src/neuro_adaptive_fusion/
│   ├── package.xml
│   ├── setup.py
│   └── neuro_adaptive_fusion/
│       ├── innovation_monitor.py # Sliding-window feature extraction
│       ├── dnn_inference.py      # TFLite inference + Eq 30
│       ├── covariance_injector.py# Mock MAVLink injection
│       └── utils/
└── launch/
    └── system.launch.py

Hardware

Component	Role	Interface
Cube Orange+	Flight Controller	MAVLink (UART)
Raspberry Pi 5	Companion Computer	USB / UART
u-blox NEO-M9N	GPS Receiver	I²C
ICM-42688-P	IMU (on Cube)	SPI (internal)

---

License

MIT