ANIMA SHEPHERD — Wave 10 WARDOG

Paper: A Multimodal Transformer for UAV Detection (Radar+RGB+IR+Audio)
ArXiv: https://arxiv.org/abs/2511.15312
Status: Initial implementation baseline (paper verified with reproducibility risks)
Focus: UAV/Drone defense for Shenzhen Robot Fair

What is implemented in this initial stage

• Paper-audit workflow with explicit red-flag reporting
• Dual-backend runtime resolution (mlx|cuda|cpu) through device.py
• Paper-aligned multimodal Transformer baseline:
  • Input projection (128 -> 256)
  • Positional encoding
  • 2-layer Transformer encoder, 4 heads
  • Mean pooling + MLP classifier
• Synthetic multimodal data pipeline (audio/rgb/ir/radar) with early-fusion tensor
• Training/evaluation loop for initial sanity checks
• YOLO26 adapter scaffold for integration with latest YOLO26 weights
• Numerical parity kernel check between MLX and Torch arithmetic paths

Quick start

# install editable package
uv pip install -e ".[dev]"

# verify paper audit snapshot
python3 -m anima_shepherd --config configs/default.toml paper-audit

# run initial synthetic training (fast sanity run)
python3 -m anima_shepherd --config configs/default.toml train-synthetic --backend auto --epochs 1 --seq-len 64

# run single forward pass on selected backend
python3 -m anima_shepherd --config configs/default.toml infer-synthetic --backend mlx --batch-size 2
python3 -m anima_shepherd --config configs/default.toml infer-synthetic --backend cpu --batch-size 2

# MLX parity kernel check
python3 -m anima_shepherd --config configs/default.toml check-parity --backend mlx --atol 1e-2

# inspect YOLO26 integration readiness
python3 -m anima_shepherd --config configs/default.toml yolo26-status

Dataset readiness check

bash scripts/download_data.sh --check
bash scripts/download_data.sh --download

The script is intentionally safe in this phase: it checks volume state and prepares acquisition notes for missing datasets without forcing downloads.

Repository layout

project_shepherd/
├── src/anima_shepherd/
│   ├── cli.py
│   ├── config.py
│   ├── data.py
│   ├── device.py
│   ├── modeling.py
│   ├── paper_audit.py
│   ├── training.py
│   └── yolo26_adapter.py
├── configs/default.toml
├── scripts/download_data.sh
├── PRD.md
├── VERIFICATION_REPORT.md
├── ARCHITECTURE.md
├── MODULE_TODO.md
└── NEXT_STEPS.md