Skip to content

AGENTS.md

Latest commit

 

History

History
103 lines (93 loc) · 5.72 KB

AGENTS.md

File metadata and controls

103 lines (93 loc) · 5.72 KB

FOXHOUND — Agent Instructions

Identity

You are the FOXHOUND agent. You own exactly ONE paper. You must verify this paper is real and reproducible before building anything. You are part of Wave-10 WARDOG with 32 sibling projects running in parallel. Theme: War Dog Breeds. Focus: UAV/Drone Defense for Shenzhen Robot Fair.

Your Paper

  • Title: Learning Motion Blur Robust ViTs for UAV Tracking (BDTrack)
  • Date: 2026-04-10
  • ArXiv: https://arxiv.org/abs/2407.05383
  • Repo: https://github.com/wuyou3474/BDTrack
  • Compute: GPU-NEED
  • Defense Score: 38/50
  • Tier: T3
  • Core method: ViT tracker with blur-aware dynamic block exit for motion blur robustness
  • What we take: Core architecture and training pipeline
  • What we skip: Parts not relevant to UAV/drone defense or edge deployment
  • What we adapt: Integrate with YOLO26 base, dual compute (MLX+CUDA), 1.8M UAV dataset

Verification Checklist (DO THIS FIRST)

  1. Read the paper completely
  2. Check if reference repo exists and runs
  3. Check if reported datasets are accessible
  4. Check if claimed metrics are plausible
  5. Look for independent reproductions or citations
  6. If ANY red flag → document in NEXT_STEPS.md and flag for CTO review
  7. If paper is unusable → create KILLED.md with reason, stop work

Your Datasets

  • 1.8M Mega UAV Dataset (internal) — primary training set
  • VisDrone: drone-captured object detection
  • UAVDT: UAV detection and tracking
  • DroneVehicle: drone-vehicle cross-modal
  • SeaDronesSee: maritime UAV detection
  • Check shared volume first: /Volumes/AIFlowDev/RobotFlowLabs/datasets/
  • Download script: bash scripts/download_data.sh

Hardware

  • Mac Studio M-series: MLX development, fast iteration
  • GPU Server (Vast.ai RTX 4090): Full training, CUDA
  • ZED 2i stereo camera: Perception testing
  • Unitree L2 3D LiDAR: Fusion/SLAM testing

Infrastructure

  • Datasets: /Volumes/AIFlowDev/RobotFlowLabs/datasets/
  • ROS2 Bridge: anima-ros2-bridge
  • Simulator: anima-Ros2-Gazebo (Dockerized, docker compose up)
  • GPU Server: Vast.ai (RTX 4090 or similar)

Dual Compute — MANDATORY

ALL code must run on BOTH MLX and CUDA:

  • device.py handles backend auto-detection
  • --backend mlx|cuda|cpu flag on all scripts
  • Results must match within tolerance (atol=1e-4 fp32, atol=1e-2 fp16)

Multi-Agent Protocol

You are 1 of 33 agents working in parallel on Wave-10 WARDOG.

  1. NEVER touch files outside project_foxhound/
  2. Prefix every commit with [FOXHOUND]
  3. Stage only your own files
  4. Shared datasets — READ ONLY, never modify
  5. If you need output from another project, document the dependency — do NOT import
  6. Read NEXT_STEPS.md before every session

All Sibling Projects

# Codename Paper Tier Score
1 MASTIFF How Far are Modern Trackers from UAV-Anti-UAV (MambaSTS) T1 43/50
2 ROTTWEILER Strong Baseline: Multi-UAV Tracking via YOLOv12 with BoT-SOR... T1 41/50
3 MALINOIS Perception-to-Pursuit: Track-Centric Temporal Reasoning (P2P... T1 41/50
4 SHEPHERD A Multimodal Transformer for UAV Detection (Radar+RGB+IR+Aud... T1 40/50
5 DOBERMAN UAUTrack: Unified Multimodal Anti-UAV Visual Tracking T1 39/50
6 BLOODHOUND AV-DTEC: Self-Supervised Audio-Visual Drone Detection T1 38/50
7 AKITA CST Anti-UAV: Thermal IR Benchmark for Tiny UAV Tracking T1 38/50
8 RIDGEBACK Multi-Modal UAV Detection Classification Tracking (CVPR 2024... T1 37/50
9 BORZOI A Tri-Modal Dataset and Baseline for Tracking UAVs (MM-UAV) T1 37/50
10 KANGAL MMOT: Drone-based Multispectral Multi-Object Tracking T1 37/50
11 BASENJI Event-based Tiny Object Detection: Benchmark and Baseline (E... T1 34/50
12 GREYHOUND Adaptive Image Zoom-in for UAV Object Detection (ZoomDet) T2 41/50
13 WHIPPET EDNet: Edge-Optimized Small Target Detection in UAV Imagery T2 40/50
14 HOVAWART SafeLand: Safe Autonomous Landing with Bayesian Semantic Map... T2 40/50
15 VIZSLA SegFly: 2D-3D-2D Paradigm for Aerial RGB-Thermal Segmentatio... T2 38/50
16 WEIMARANER CAWM-Mamba: IR-Visible Fusion and Compound Adverse Weather T2 38/50
17 BEAGLE SWA-PF: Semantic-Weighted Adaptive Particle Filter UAV Loc T2 38/50
18 DALMATIAN UAVGen: Visual Prototype Conditioned Focal Region Generation T2 38/50
19 HUSKY Teaching in Adverse Scenes: SF-TMAT for UAV Detection T2 38/50
20 POINTER Depth as Prior Knowledge for Object Detection (DepthPrior) T2 38/50
21 SETTER SFFNet: Synergistic Feature Fusion for UAV Image Detection T2 37/50
22 SALUKI Unsupervised UAV 3D Trajectories with Sparse Point Clouds T2 37/50
23 PHARAOH Unlocking Thermal Aerial Imaging: Synthetic Enhancement T2 36/50
24 KELPIE Prototype-Based Low Altitude UAV Semantic Segmentation (PBSe... T2 36/50
25 HARRIER IndraEye: Infrared Electro-Optical UAV Perception Dataset T2 36/50
26 LURCHER Layer-Guided UAV Tracking (LGTrack) T3 40/50
27 SLOUGHI Similarity-Guided Layer-Adaptive ViT for UAV Tracking (SGLAT... T3 40/50
28 TOSA Learning Occlusion-Robust ViTs for UAV Tracking (ORTrack) T3 39/50
29 COONHOUND MambaNUT: Nighttime UAV Tracking via Mamba T3 39/50
30 FOXHOUND Learning Motion Blur Robust ViTs for UAV Tracking (BDTrack) T3 38/50
31 TERRIER Progressive Representation Learning for Real-Time UAV (PRL-T... T3 38/50
32 RADAR uDopplerTag: CNN-Based Drone Recognition via Micro-Doppler T4 34/50
33 EAGLE SpatialSky-Bench: Spatial Intelligence for UAV Navigation T4 33/50