ANIMA HARRIER

Paper: IndraEye: Infrared Electro-Optical UAV-based Perception Dataset for Robust Downstream Tasks ArXiv: https://arxiv.org/abs/2410.20953 Public repo: https://github.com/airl-iisc/IndraEye Wave: 10 WARDOG — UAV/Drone Defense for Shenzhen Robot Fair

What HARRIER Ships

HARRIER turns IndraEye into an ANIMA-ready UAV perception module with a YOLO26-first detection stack, mandatory dual-backend support (CUDA + MLX), and the full ANIMA serving / ROS2 / export story.

• Data — paper-faithful class list, Table III split counts, Ultralytics YAML renderer, synthetic fixtures for CI.
• Training — backend-aware HarrierTrainer wrapping YOLO26 (MuSGD, seed=42, patience=10), artifact layout rooted at /mnt/artifacts-datai, --dry-run planning mode.
• Inference — modality-aware HarrierPredictor, single-shot CLI, latency benchmark.
• Export — mandatory chain pt -> safetensors -> ONNX -> TRT fp16 -> TRT fp32 codified as data.
• Serving — FastAPI /health /ready /predict, three Dockerfiles (cuda, mlx, serve), GPU reservation compose + HEALTHCHECK.
• ROS2 — EO + IR sensor_msgs/Image in, Detection2DArray + diagnostics out, launch file, pure-Python handler.
• Evaluation — paper Table VIII metric registry with tolerance checks, markdown report writer.

Status

• Paper verified, YOLO26 rebase documented.
• 46 unit tests green (stdlib only, no torch / ultralytics / fastapi required for CI).
• 7 PRDs built, tested, committed.
• Training is BLOCKED until NIGHTHAWK mega-dataset build clears GPUs 2-7 and the IndraEye download gate opens. The code is ready to run the second both unblock.

Quick Start

# Inspect module state — no heavy deps needed
python3 -m anima_harrier info
python3 scripts/prepare_data.py --audit-only

# Plan a training run (no GPU touched)
python3 scripts/train.py \
    --experiment eo_daynight_to_eo_day \
    --backend cpu \
    --dry-run

# Run the full test suite
python3 -m pytest tests/ -q

Project Layout

• PRD.md — module-level build brief
• NEXT_STEPS.md — execution ledger + MVP score
• ASSETS.md — datasets, weights, paper metrics
• prds/ — 7-PRD execution plan
• tasks/ — granular task queue
• src/anima_harrier/ — package source
• scripts/ — CLI entrypoints (train, predict, export, benchmark, serve, prepare_data)
• docker/ — CUDA + MLX + serve Dockerfiles
• ros/ — ROS2 node + launch file

License

Research planning and integration scaffold. Check upstream dataset and model licenses before training or deployment.