ANIMA BEAGLE — Wave 10 WARDOG

Paper: SWA-PF: Semantic-Weighted Adaptive Particle Filter for Memory-Efficient 4-DoF UAV Localization in GNSS-Denied Environments ArXiv: https://arxiv.org/abs/2509.13795 Reference Repo: https://github.com/YuanJiayuuu/SWA-PF Defense Score: 38/50 | Tier: T2 Focus: GNSS-denied UAV localization for UAV/drone defense in Shenzhen demo stacks

Verification Snapshot

• Paper PDF is present locally and matches the public arXiv paper.
• Public code exists and compiles, but it is research-grade and pinned to an old PyTorch 1.10 + CUDA 11.3 environment.
• Public dataset access is partial and gated. The shared volume currently exposes VisDrone, but not a ready-to-train MAFS mirror.
• Independent reproductions were not found during the initial verification pass.
• Verdict: CONDITIONAL GO for ANIMA baseline work. Reproduction is feasible, but exact paper replication remains blocked on dataset access and missing pretrained artifacts.

What BEAGLE Builds

BEAGLE implements the paper's core method as a clean ANIMA baseline:

• semantic map preprocessing
• semantic-weighted adaptive particle filtering
• 4-DoF pose state propagation and weighting
• dual backend support: mlx | cuda | cpu
• future adaptation hook for YOLO26-driven defense stacks

YOLO26 is not the paper's core model. It is reserved for later adaptation and fusion work, not for the first-paper reproduction target.

Quick Start

uv pip install -e ".[dev]"
python3 -m anima_beagle inspect-config
python3 -m anima_beagle dataset-status
python3 -m anima_beagle synthetic-demo --backend auto

Project Structure

project_beagle/
├── ASSETS.md
├── PRD.md
├── prds/
├── tasks/
├── src/anima_beagle/
├── tests/
├── configs/
├── scripts/
├── papers/
└── repositories/SWA-PF/

Reference Notes

• Satellite semantics in the paper use a small VGG U-Net.
• UAV semantics use SegFormer-B0.
• Localization is driven by semantic-distance maps, a center distance field, and adaptive particle resampling/clustering.
• The paper reports RMSE 6.57 m, Recall@10 97.37%, and ~7 s fitting time on MAFS-10.

Status

The current codebase is a phase-1 ANIMA scaffold with a paper-faithful particle filter core and synthetic verification path. Real-dataset reproduction remains dependent on MAFS access and later semantic model weight integration.