Skip to content

ArgusAiTR — Documentation#

Aided Target Recognition (AiTR) overlay for Remote Weapon Stations. Addressing DIU CSO — Endpoint Accuracy 2: CUAS Close-In Kinetic Defeat Enhancement, Part 1.

ArgusAiTR is a non-intrusive AiTR module that taps an RWS (CROWS) video output, runs passive detection / classification / multi-object tracking / monocular ranging on commodity edge compute, and renders overlays + slew-to-cue back to the operator. Mandatory human-in-the-loop on every engagement decision; graceful fallback to standard RWS operation if the AiTR pipeline degrades.

Scope of this submission

This documentation accompanies the 48-hour PoC artifact submitted with Argus Defense's DIU initial application. The pipeline is real (ingest → perturb → stabilize → detect → track → range → classify → HITL HMI → audit). Real CROWS hardware integration, fine-tuned UAS detectors, supplementary EO/IR sensors, and live-fire validation are scoped as post-award work — see the Post-award roadmap.

For reviewers#

Start here:

  1. Technical brief — the full submission narrative (~8 pages). Cover, executive summary, problem understanding, technical approach, demonstration, AI ethics, IP/licensing, team/scaling, schedule, compliance checklist.
  2. One-pager — condensed summary; Part-1 callout; what's submitted; CSO compliance posture.
  3. Architecture — data flow, module contracts, threading model, failure modes.
  4. DoD AI Ethics compliance — five-principle table mapped to specific code locations.

CSO requirement coverage#

Requirement Where addressed
Passive detection / classification / tracking src/argus_aitr/detect/, track/, classify/
Monocular ranging src/argus_aitr/range/monocular.py
Adverse conditions (jitter, muzzle flash) src/argus_aitr/stabilize/{stabilizer,perturb}.py
Cluttered backgrounds Detection confidence floor + kinematic threat heuristic + audit log
HITL hard gate src/argus_aitr/hmi/server.py (POST /op/engage_request requires modal confirm)
Graceful fallback Pipeline health monitor → HMI banner; /aitr/disable for manual cutover
Open system architecture Pydantic contracts + ABCs at every stage boundary
DoD AI Ethical Principles docs/ai-ethics.md (table mapped to code locations)

What the demo shows#

  1. Live AiTR loop — boxes on detected targets, threat-colored, with class / range / confidence / age. Click any target for HITL action buttons (Confirm, Engage, Dismiss).
  2. Adverse-conditions toggle — switch the perturbation profile from clean to engagement (jitter + muzzle flash) or severe (adds dropout). The tracker holds through the simulated disturbance. Directly addresses the CSO clause on "shake from weapon firing, muzzle flash, and high frequency jitter."
  3. HITL hard gate — every engagement click triggers a confirmation modal and is recorded as an operator_action in the audit log. The mock CROWS adapter logs the slew command instead of sending it.
  4. Failsafe — click "Disable AiTR" or watch FPS auto-drop the system to degraded. The operator console shows a prominent red banner and the operator continues with the standard RWS view.
  5. Audit panel — last 5 operator actions surfaced in the sidebar; full audit JSONL written to audit/argus.jsonl.

Status#
Tests 111 passing on Apple Silicon Python 3.11
LOC ~3 000 source Python + ~300 frontend + 9 documents
License Apache-2.0 (Argus-authored); see Licensing for OSS deps
Repository https://github.com/PoggyBobby/argus-aitr (private; access on request)

Contact#

Argus Defense — production@argusdefense.us