- Codename:
LURCHER - Focus: UAV/Drone tracking for Shenzhen Robot Fair defense scenarios
- Paper: Layer-Guided UAV Tracking (LGTrack), arXiv:2602.13636
- Reference code: https://github.com/XiaoMoc/LGTrack
- Current phase: Initial autopilot bootstrap completed
Build a production-ready ANIMA module that reproduces and adapts LGTrack’s layer-guided ViT tracking for UAV defense workloads, with mandatory dual backend support (mlx, cuda, cpu) and YOLO26-compatible integration points.
- Scope only inside
project_lurcher/. - All runtime entrypoints must accept
--backend mlx|cuda|cpu|auto. - Cross-backend numerical parity targets:
- fp32:
atol <= 1e-4 - fp16:
atol <= 1e-2
- fp32:
- If detector hooks are needed, default family is YOLO26.
- Shared datasets are read-only.
- Paper PDF exists locally and was fully read (
papers/2602.13636.pdf). - Reference implementation exists and is non-empty (
repositories/LGTrack). - Core claims and training recipe were extracted from paper sections §3.x and §4.x.
- Core reported benchmark points captured:
- UAVDT: Prec 82.8, Succ 60.4, 258.7 FPS (GPU)
- CPU speed reference: 59.4 FPS
- External reproductions/citations are still limited (new paper risk).
- Upstream repo lacks explicit license file.
- Public UAVDT URL reliability was inconsistent during checks.
- Upstream codebase targets older CUDA/PyTorch versions.
Paper is usable for implementation and adaptation. Continue execution with CTO visibility on risks.
- Implement paper-grounded components:
- SGLA dynamic deep-layer selection
- GGCA attention refinement
- ORR occlusion-robust consistency loss
- Build train/infer scaffolds in
src/anima_lurcher/. - Generate full PRD/task suite and execution index.
- Provide dataset preflight tooling and reproducibility checklist.
- Full scale multi-day training on 1.8M internal data.
- Final ROS2/Gazebo deployment hardening.
- TensorRT packaging and full edge benchmark matrix.
- ANIMA perception engineers reproducing paper baselines.
- Wave-10 integrators composing multi-module UAV defense stack.
- Shenzhen demo operators requiring low-latency UAV tracking under occlusion.
- Central TOML config exists and loads deterministically.
- Paper hyperparameters are represented explicitly in config.
- A testable LGTrack-like model exists with:
- Transformer token flow
- SGLA selector and labels
- GGCA feature refinement
- ORR training-time consistency branch
- CLI inference smoke path runs on CPU by default.
- YOLO26 loader hook is present and lazy.
- CLI dry-run training loop executes end-to-end.
- Includes bbox loss, ORR loss, and SGLA selector supervision.
- Unit tests cover config/device/model shape assumptions.
- All tests run from project root with
pytest -q.
- Full
ASSETS.md,prds/, andtasks/suite exists. NEXT_STEPS.mdandMODULE_TODO.mdreflect current state.
- Deterministic behavior for smoke tests.
- No hard dependency on MLX/CUDA for CPU smoke tests.
- Backends resolved in priority order when
autois used. - Codebase remains modular for later API/ROS2 integration.
- Input template/search images.
- Patch embedding + Transformer blocks.
- SGLA predicts optimal deep layer activation after saturation point (
l*=8). - Search tokens reshaped to feature map and refined by GGCA.
- Score/box heads produce prediction maps.
- Training mode computes ORR consistency under Cox-style masking.
src/anima_lurcher/device.py— backend resolver and tolerance helperssrc/anima_lurcher/config.py— typed config loadingsrc/anima_lurcher/model/sgla.py— selector and similarity labelssrc/anima_lurcher/model/attention.py— GGCA implementationsrc/anima_lurcher/model/occlusion.py— ORR masking/loss helperssrc/anima_lurcher/model/lgtrack.py— assembled LGTrackLite modelsrc/anima_lurcher/infer.py— inference smoke entrypointsrc/anima_lurcher/train.py— training dry-run entrypointsrc/anima_lurcher/yolo26.py— YOLO26 integration hook
- Shared volume first:
/Volumes/AIFlowDev/RobotFlowLabs/datasets/ - Preflight script:
bash scripts/download_data.sh - Current observed state:
- VisDrone staging folder found
- UAVDT/DroneVehicle/SeaDronesSee presence not yet fully validated locally
- 1.8M internal dataset not yet enumerated
- Paper/repo verified
- Initial code scaffold
- Unit tests green
- PRD/task suite generated
- Wire real dataset loaders
- Reproduce close-to-paper metrics on UAVDT/UAV123/DTB70
- Document deviations and reasons
- Integrate YOLO26 tracks and internal 1.8M data
- Run dual backend parity checks
- Produce updated benchmark report
- API serving + Docker profile
- ROS2 bridge and launch artifacts
- Production export and monitoring
- Paper and repo read/verified with explicit risk notes.
- Initial LGTrack-based code exists and runs smoke paths.
- Tests pass locally (
5 passed). - Full 7-PRD suite exists under
prds/. - Granular task suite exists under
tasks/with build order index. - YOLO26 default is represented in configuration and integration hook.
- Internal datasets and model storage mounts.
- Optional GPU server for heavy training.
- Sibling Wave-10 YOLO26 fine-tune output (if consumed directly).
- Reproduction drift vs paper numbers.
- Mitigation: keep phase-2 benchmark report with exact settings and deltas.
- Dataset gaps or format inconsistencies.
- Mitigation: data adapter layer + schema checks before training.
- Backend divergence (MLX vs CUDA).
- Mitigation: parity tests with tolerance thresholds in CI.
- Full documentation scaffold for
/anima-autopilotexecution. - Executable initial code package for LGTrack adaptation.
- Verification ledger updates and actionable next steps.