ArgusAiTR — one-pager#
DIU CSO — Endpoint Accuracy 2: CUAS Close-In Kinetic Defeat Enhancement Addressing Part 1 only. Aided Target Recognition for CROWS-class RWS against Group 1/2 UAS at 100–600 m, with mandatory human-in-the-loop and graceful fallback.
Vendor: Argus Defense · production@argusdefense.us · CAGE/SAM registration in progress
The system#
A non-intrusive AiTR overlay. ArgusAiTR taps the CROWS video output, runs passive detection / multi-object tracking / monocular ranging / kinematic threat classification on commodity edge compute, and renders overlays + slew-to-cue back to the operator. CROWS firmware is unchanged. If AiTR drops, the operator retains full standard RWS capability — verifiably, via a single click or automatic FPS/error-based degradation.
Submitted artifact#
- Public GitHub repository with the full pipeline (ingest → stabilize → detect → track → range → classify → operator HMI → audit). One-command launcher; 111 tests pass on Apple Silicon Python 3.11.
- Hosted demo on Hugging Face Space (free CPU tier) for reviewers to poke without install.
- 3–5 min recorded walkthrough demonstrating the operator workflow, the adverse-conditions perturbation toggle, the failsafe banner, and the audit log.
What the demo shows#
- Live AiTR loop with threat-colored boxes, class / range / confidence, and HITL action buttons.
- Adverse-conditions toggle — switching to
engagementinjects jitter + muzzle flash + vibration into the live feed; the tracker holds through it. Directly addresses the CSO clause on "shake from weapon firing, muzzle flash, and high frequency jitter." - HITL hard gate — every engage click triggers a confirmation modal and is audited; the mock CROWS adapter logs the slew command instead of sending it. The system never issues a slew command without an operator action.
- Failsafe — clicking "Disable AiTR" or auto-degradation flips the HMI banner to red; the operator continues with standard RWS view.
CSO compliance posture#
| Requirement | Status |
|---|---|
| Open system architecture | ✓ Pydantic contracts + ABCs at every stage boundary |
| HITL hard gate | ✓ Architecturally enforced; auditable |
| Graceful fallback | ✓ Auto-degrade on FPS/error + manual /aitr/disable |
| Adverse conditions | ✓ Perturbation simulator with 4 named profiles, switchable from HMI |
| DoD AI Ethical Principles | ✓ Compliance table in docs/ai-ethics.md mapped to code |
| Clear IP rights | ✓ Apache-2.0 Argus code; AGPL Ultralytics has documented migration to RT-DETR (Apache-2.0) or commercial license post-award |
| Not a research-only org / reseller / integrator | ✓ Argus is a product company; this artifact is the product |
Honest baseline + scaling plan#
Argus today is one person + Claude Code as development partner. The 48-hour artifact demonstrates the architectural judgment and execution speed of that pairing. Post-award:
- Month 1: Hire ML + systems engineers. Engage Kongsberg US (CROWS OEM) for surrogate access. Source supplementary EO + cooled MWIR (FLIR / Teledyne). Port pipeline to Jetson AGX Orin.
- Month 2: Acquire Argus-licensed UAS imagery; fine-tune RT-DETR. Upgrade tracker for dense scenes; add platform-motion triangulation. Range testing to 600 m envelope.
- Month 3: Government range scheduled. Operator rehearsals. Live-fire UAS engagement demonstration.
Partnership pipeline: Kongsberg US, NSWC / Army C5ISR Center, FLIR / Teledyne, Edmund / MOOG, NVIDIA. Multiple paths exist for every critical resource so no single partnership is a single point of failure.
Why Argus#
The reviewer who reads docs/architecture.md, runs uv run python scripts/run_demo.py, and clicks the perturbation toggle has, in ten minutes, the full evidence that the architectural decisions are correct, the engineering execution is real, and the team's posture is honest. The submitted artifact is the calibration. The roadmap is the plan to deliver against it.
Repository: https://github.com/PoggyBobby/argus-aitr
Full technical brief: docs/proposal/argus-cso-part1-brief.md