Skip to content

NEXT_STEPS.md

Latest commit

 

History

History
111 lines (93 loc) · 4.92 KB

NEXT_STEPS.md

File metadata and controls

111 lines (93 loc) · 4.92 KB
Raw

BEAGLE — Execution Ledger

Last updated: 2026-04-11 MVP Readiness: 65%

Resume rule: Read this file COMPLETELY before writing any code. This project covers exactly ONE paper: SWA-PF: Semantic-Weighted Adaptive Particle Filter for Memory-Efficient 4-DoF UAV Localization in GNSS-Denied Environments.

1. Working Rules

  • Work only inside project_beagle/
  • This wave has 33 parallel projects, 33 papers, 33 agents
  • Prefix every commit with [BEAGLE]
  • Stage only project_beagle/ files
  • Do NOT start dataset downloads until download gate is opened
  • VERIFY THE PAPER BEFORE BUILDING ANYTHING
  • YOLO26 is adaptation-only unless a detector-specific submodule is added

2. The Paper

3. Datasets

Required for this paper

Dataset Size URL Format Phase Needed
1.8M Mega UAV ~1.8M imgs Internal YOLO format Phase 2-3
VisDrone ~10K imgs github.com/VisDrone COCO Phase 2
UAVDT ~80K frames uavdt.github.io MOT Phase 2

Check shared volume first

/Volumes/AIFlowDev/RobotFlowLabs/datasets/ Many datasets are shared across projects. Check before downloading.

Download

bash scripts/download_data.sh

4. Hardware

  • Mac Studio M-series: MLX development
  • GPU Server (Vast.ai): CUDA training
  • ZED 2i: Stereo perception
  • Unitree L2 LiDAR: 3D fusion

5. Infrastructure

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

6. Active Execution Backlog

Done

  • Read paper completely
  • Check reference repo (clone, inspect, syntax compile)
  • Verify reported metrics are plausible
  • Check for independent reproductions
  • Identify required datasets
  • Check shared dataset volume for existing data
  • Rewrite download_data.sh as a gate-aware inventory helper
  • Design module architecture (update PRD.md)
  • Generate full PRD suite and tasks
  • Implement Phase 1: scaffold + synthetic tests
  • Implement segmentation models (SegFormer-B0 + U-Net VGG)
  • Implement training pipeline (two-phase, early stopping, checkpointing)
  • Implement evaluation harness (RMSE, Recall@K, mIoU)
  • Implement export pipeline (pth, safetensors, ONNX, TRT)
  • Implement FastAPI serving endpoint
  • Create dual Docker setup (CUDA + MLX)
  • Create anima_module.yaml manifest
  • Create PIPELINE_MAP.md
  • Code review and fixes

TODO

  • Acquire MAFS dataset or approved substitute
  • Run training on GPU (freeze 50ep + finetune 200ep)
  • Evaluate on test split, generate TRAINING_REPORT.md
  • Run full export pipeline on trained model
  • Push trained model to HuggingFace
  • ROS2 node integration
  • Simulator bridge

6a. Verification Notes

  • Paper is real and locally available as papers/2509.13795.pdf.
  • Public repo was cloned into repositories/SWA-PF.
  • Public repo compiles with python3 -m py_compile, but uses a legacy environment and hard-coded dataset paths.
  • Shared dataset volume currently exposes VisDrone staging only. MAFS is not present locally.
  • Public MAFS access uses a Baidu share link with passcode, so exact paper reproduction remains gated.
  • No independent reproduction was found during this initial pass.
  • CTO review should focus on whether to request MAFS/SemanticMAFS ingress or proceed directly with ANIMA adaptation on internal UAV data.

Blocking

  • MAFS dataset not available (Baidu-gated). Need CTO approval for substitute or ingress.
  • GPUs 2-7 busy with NIGHTHAWK mega dataset generation. Use GPU 0 or 1 when ready.
  • Pretrained SegFormer-B0 and U-Net weights not yet downloaded.

Downloads Needed

  • SegFormer-B0 ADE20K pretrained weights
  • MAFS dataset (Baidu link, requires passcode from paper authors)

7. Session Log

Date Agent What happened Next
2026-04-10 ANIMA Research Agent Project scaffolded, mandatory docs created Read paper, verify, clone repo
2026-04-10 Codex / BEAGLE Verified paper and public repo, documented dataset/repro blockers, generated ANIMA PRDs/tasks, implemented phase-1 SWA-PF core scaffold Obtain MAFS assets or approved substitute
2026-04-11 BEAGLE Autopilot Full build: segmentation models (SegFormer-B0 + U-Net VGG), training pipeline (two-phase with early stopping), evaluation harness, export pipeline (5 formats), FastAPI serving, dual Docker (CUDA+MLX), integration files. Code review done. 5 commits. Ready for training when GPUs free. Acquire dataset, train on GPU 0/1