I’m building a proof-of-concept drone interceptor that relies on the Ardupilot flight stack and an onboard AI module. The goal is clear: once airborne, the aircraft must identify other drones in its vicinity, lock on, then maintain a stable track and follow pattern—all without human stick time. Core requirements • Ardupilot will handle all flight control, so your work focuses on integrating the AI pipeline with its GUIDED modes and mavlink commands. • The AI must deliver reliable, real-time object recognition that flags and classifies “drone” targets from the video feed, pushing bounding-box coordinates back to the flight controller at 20 Hz or better. • On receipt of those coordinates, the autopilot should autonomously manoeuvre to keep the target centred, holding a safe standoff distance that I can configure via a parameter file. Deliverables 1. Clean, well-documented code for the vision stack (Python/C++, OpenCV, TensorFlow or similar) plus a companion Ardupilot lua or mavlink-based script that interprets vision output and issues NAV commands. 2. A step-by-step setup guide covering hardware connections, firmware parameters, and any custom GStreamer pipelines required for the video link. 3. Short demo footage or SITL log proving that the system detects a drone-class object and keeps it in frame for at least 60 seconds. Acceptance will be based on flawless compilation, smooth SITL tests, and a final field test on my quad running latest Ardupilot stable. If you’ve previously tied AI vision into PX4 or Ardupilot, let’s chat—I’m ready to move fast.