detections-semantic/_docs/_autopilot_state.md

Autopilot State

Current Step

step: 3 name: Decompose status: in_progress sub_step: 1 — Bootstrap Structure Plan

Completed Steps

Step	SubStep	Name	Completed	Key Outcome
0	4/4	Problem	2026-03-19	Semantic detection for camouflaged positions via two-level scan, 3 submodules
1	4/4	Research (Mode A)	2026-03-19	Draft01: three-tier architecture (YOLOE-26 + CNN + VLM)
1	8/8	Research (Mode B R1)	2026-03-19	Draft02: 11 weak points addressed — thermal, robustness, degradation, VLM availability hedge
1	8/8	Research (Mode B R2)	2026-03-19	Draft03: 11 more weak points — vLLM replaced with NanoLLM, dual backbone, NVMe mandatory, CRC UART, recording/logging, power management
2	6/6	Plan	2026-03-20	6 components + 2 helpers, 93 tests across 6 specs, 8 epics (42-68 pts), 96% AC coverage, FINAL_report.md complete

Key Decisions

• Three-tier architecture with graceful degradation
• YOLOE backbone configurable: benchmark YOLO11 vs YOLO26 before committing
• FP16 TRT only for initial deployment (INT8 unstable on Jetson)
• NanoLLM replaces vLLM as VLM runtime (vLLM unstable on Jetson)
• VILA1.5-3B primary VLM via NanoLLM; UAV-VL-R1 via llama.cpp if GGUF available
• NVMe SSD mandatory (SD card corruption documented)
• UART integrity: check ViewLink spec first for native checksum, then add CRC-16 if needed
• Detection logger + frame recorder for post-flight review and training data
• Power monitoring via INA sensors + load shedding
• Ruggedized carrier board (MILBOX-ORNX or similar)
• V1 path tracing: minimal heuristic, no CNN (ship fast, validate)
• V2 CNN removed from plan — heuristic + VLM sufficient; CNN adds unnecessary complexity
• Version pinning: Ultralytics, JetPack, NanoLLM
• E2E test coverage of 76% accepted
• Data model simplified: no traditional DB, only runtime structs + persistent flat files (NVMe)
• Architecture simplified: 5 system flows, inline health checks, capability flags, 2 environments
• ScanController: Behavior Tree (py_trees 2.4.0) — ADR-008
• Dynamic Search Scenarios — data-driven YAML configs with 4 investigation types (path_follow, cluster_follow, area_sweep, zoom_classify)
• GIL is not a concern — all compute-heavy ops release GIL; VLM in separate process; single-threaded by design
• Tier2PathTracer generalized to Tier2SpatialAnalyzer — supports mask tracing (footpaths) and cluster tracing (AA networks, vehicle groups) via unified SpatialAnalysisResult/Waypoint output
• Jira epics created: AZ-130 (Bootstrap), AZ-131 (Tier1), AZ-132 (Tier2), AZ-133 (VLM), AZ-134 (Gimbal), AZ-135 (Output), AZ-136 (ScanController), AZ-137 (Integration Tests)

Last Session

date: 2026-03-20 ended_at: Step 2 Plan — SubStep 6/6 complete reason: Plan step completed, auto-chaining to Step 3 Decompose notes: Steps 5 (Test Specifications) and 6 (Jira Epics) completed. 93 tests written across 6 component test specs (50 integration, 14 performance, 9 security, 20 acceptance). 8 epics documented in epics.md (Jira MCP auth skipped). FINAL_report.md written. 96% AC coverage (27/28 — AC-28 training dataset is data annotation scope, not runtime).

Blockers

• R05 (Critical): Seasonal generalization — mitigated by phased rollout (winter first)