ANIMA HOVAWART — Wave 10 WARDOG

Paper: SafeLand: Safe Autonomous Landing with Bayesian Semantic Mapping ArXiv: https://arxiv.org/abs/2603.17430 GitHub: https://github.com/markus-42/SafeLand Defense Score: 40/50 | Tier: T2 Wave: 10 — WARDOG (War Dog Breeds) Focus: UAV/Drone Defense for Shenzhen Robot Fair

Overview

HOVAWART implements the SafeLand landing stack for ANIMA with a pragmatic Wave-10 adaptation:

• SafeLand core: semantic segmentation -> metric projection -> Bayesian semantic map -> landing spot selection -> behavior-tree landing control
• HOVAWART adaptation: YOLO26 human veto path, dual compute support (mlx + cuda), and integration hooks for the Wave-10 UAV stack

Published result: 70.22% mIoU segmentation, 95% end-to-end landing success, zero false negatives for human detection, sub-second obstacle response

Verification Status

The paper is real and the preprint is internally coherent, but reproducibility is only partial today.

• Paper: available locally in papers/2603.17430.pdf
• Reference repo: cloned under references/SafeLand, but currently contains only a README and image with “Code coming soon”
• Datasets: visdrone is present on the shared volume; the rest of the paper dataset bundle is not yet clearly staged
• Verdict: usable for initial system implementation, not yet reproducible for paper-faithful training

See PRD.md, ASSETS.md, prds/README.md, and tasks/INDEX.md for the full build manual.

Quick Start

# Install dependencies
uv pip install -e ".[dev]"

# Run a synthetic SafeLand pipeline pass
python -m anima_hovawart

# Explicit backend override
ANIMA_BACKEND=mlx python -m anima_hovawart
ANIMA_BACKEND=cuda python -m anima_hovawart

Project Structure

project_hovawart/
├── src/anima_hovawart/   # Source code
├── tests/                   # Unit tests
├── configs/                 # Configuration files
├── scripts/                 # Utility scripts
├── papers/                  # Paper PDF
├── prds/                    # ANIMA PRD suite
├── tasks/                   # Task breakdown for execution
├── docker/                  # Docker setup
├── CLAUDE.md               # Agent instructions
├── PRD.md                  # Production requirements
├── ASSETS.md               # Asset and dataset manifest
├── NEXT_STEPS.md           # Execution ledger
└── MODULE_TODO.md          # Implementation checklist

Dual Compute

All runtime code is written behind a backend abstraction that resolves to mlx, cuda, or cpu.

The initial scaffold includes:

• backend detection in device.py
• SafeLand Bayesian map and landing candidate logic
• a synthetic segmenter for local smoke testing
• a YOLO26 human-prior adapter stub for the Wave-10 deployment path

Current Scope

Implemented now:

• paper-grounded pipeline scaffold
• config loader and type models
• Bayesian semantic map update with person-safe handling
• metric landing spot selection via distance transform
• behavior-tree style landing decision state machine
• PRD suite and execution tasks

Still blocked on external deliverables:

• author code release
• published SafeLand weights / TensorRT artifacts
• full dataset bundle availability
• real ROS2 and Docker serving layers

License

Research use only. See paper for original license terms.