Skip to content

PRD.md

Latest commit

 

History

History
135 lines (114 loc) · 5.61 KB

PRD.md

File metadata and controls

135 lines (114 loc) · 5.61 KB
Raw

RADAR PRD - uDopplerTag Bootstrap and Adaptation

Status: Phase-1 Bootstrap In Progress Version: 1.0 Date: 2026-04-10 Module: RADAR (Wave-10 WARDOG #32) Paper: uDopplerTag: CNN-Based Drone Recognition via Cooperative Micro-Doppler Tagging ArXiv: https://arxiv.org/abs/2601.08042

1. Objective

Build a production-oriented ANIMA module that reproduces the paper's core micro-Doppler CNN pipeline and adapts it for UAV defense operations with dual-compute support (mlx, cuda, cpu) and downstream YOLO26 integration.

2. Verification Verdict

Verification state: VERIFIED-WITH-RISKS (not killed)

What was verified:

  • Paper exists on arXiv as 2601.08042v1 (published 2026-01-12).
  • Claimed architecture and training recipe are explicit enough to implement:
    • 3 Conv2D blocks (32/64/128 filters, 3x3, ReLU)
    • MaxPool 2x2 after each block
    • Dropout p=0.5
    • FC 128 + softmax classifier
    • Cross-entropy + Adam (lr=0.001), batch size 32
  • Claimed data protocol is explicit enough for dataset shaping:
    • Indoor windows from 10 s / 50,001-sample records, 1.54 s windows
    • Outdoor windows from 15 s / 90,001-sample records, 2 s windows, 400x59 spectrogram input

Risk flags (documented for CTO review):

  • No public reference implementation repo released.
  • No public released raw dataset from paper (indoor/outdoor tagged radar captures are private).
  • No independent reproductions found yet; citation signal currently zero in OpenAlex.
  • Reported long-range feasibility is extrapolated from SNR behavior, not directly field-demonstrated in the paper.

Decision:

  • Proceed with architecture-faithful implementation and reproducible synthetic/adapter benchmarks.
  • Gate paper-metric claims until internal radar-equivalent data or author data access is available.

3. Scope

In scope now:

  • Complete planning package (ASSETS.md, prds/, tasks/).
  • Core code scaffold with paper-aligned model, preprocessing, training/eval entrypoints.
  • Dual-compute backend handling and parity harness.
  • Unit tests for config, preprocessing, model shape, backend selection.

Out of scope now:

  • True paper-data reproduction (dataset unavailable publicly).
  • End-to-end ROS2 and deployment hardening (defined in later PRDs/tasks).
  • Full YOLO26 fused runtime (integration hooks only in this phase).

4. Paper-Derived Technical Baseline

From paper Sections II-IV and appendices:

  • Radar frequency context: 3.35 GHz CW system.
  • Input representation: STFT-derived spectrogram matrices from I/Q radar streams.
  • Indoor class count used in paper: 43.
  • Outdoor class count used in paper: 7 selected robust codes.
  • Train/val/test split (outdoor protocol): 80/10/10.
  • Indoor robustness reporting via synthetic AWGN on held-out windows.
  • Indicative results:
    • About 99% accuracy above ~9 dB indoor synthetic-SNR evaluation.
    • Useful outdoor robustness reported down to 7 dB for selected code subset.

5. ANIMA Adaptation

Target adaptation for Wave-10:

  • Keep paper radar-CNN branch as anima_radar core classifier.
  • Add YOLO26 pathway as external perception prior for fused anti-UAV stack:
    • YOLO26 provides object-level hypotheses (bearing/box track context).
    • uDopplerTag branch provides identity-level class likelihood from radar signatures.
  • Long-term fusion target:
    • score_fused = f(score_yolo26, score_microdoppler, track_stability, geometry_consistency)

6. Architecture (Current + Planned)

Current phase architecture:

  1. Data ingestion (RadarPreprocessor)
    • Segment I/Q streams by indoor/outdoor windows.
    • STFT -> magnitude dB map -> resize to paper-compatible tensor shape.
  2. Model
    • Torch implementation (NCHW)
    • MLX implementation (NHWC)
  3. Backend layer
    • Auto-detect mlx/cuda/cpu
    • Explicit --backend override on CLIs
  4. Evaluation
    • Synthetic SNR sweep utilities
    • Confusion-ready outputs

Planned extension:

  • FastAPI serving + ROS2 bridge + simulator adapters + export pipeline.

7. Datasets and Assets

Primary operational datasets for ANIMA adaptation:

  • Internal 1.8M Mega UAV dataset (path still pending explicit mount mapping)
  • VisDrone (found in shared staging volume)
  • UAVDT (not found yet in shared staging)
  • DroneVehicle (not found yet)
  • SeaDronesSee (not found yet)

Paper-native data status:

  • Not publicly available.

8. Success Criteria

Phase-1 done when:

  • Planning package complete (this PRD + ASSETS.md + 7 PRDs + tasks).
  • Core model forward pass works on Torch and MLX backends.
  • Backend parity harness exists and reports numerical deltas.
  • Synthetic preprocessing and SNR-eval smoke tests pass.
  • NEXT_STEPS.md updated with verification evidence and red flags.

Program-level target (later phases):

  • Reproduce paper-like behavior on internal surrogate radar/tag data.
  • Integrate with YOLO26 prior branch and ROS2 pipeline.
  • Meet deployment latency targets on Mac Studio and RTX 4090.

9. Delivery Plan

  • PRD-01 Foundation and Data
  • PRD-02 Core Model
  • PRD-03 Inference Pipeline
  • PRD-04 Evaluation and Benchmarks
  • PRD-05 API and Docker
  • PRD-06 ROS2 Integration
  • PRD-07 Production Hardening

All details and file-level build order are in prds/README.md and tasks/INDEX.md.

10. Constraints and Guardrails

  • Work only inside project_radar/.
  • Stage only RADAR files.
  • Do not mutate shared dataset volumes.
  • Keep dual-compute support mandatory.
  • Record every red flag and unresolved dependency in NEXT_STEPS.md.

11. External Verification Links