Assistant Executive Engineer (Robotics Software & Autonomy) at NRTC | MS AI student at PAF-IAST
Rawalpindi, Pakistan • LinkedIn • Email • Portfolio Website

I work on robot perception and autonomy for UAVs: detection and tracking on edge hardware, simulation before flight (NVIDIA Isaac Sim, Gazebo, PX4), and getting useful information across degraded wireless links.

A. Naseer, A. Khalid (HITEC). First author, submitted to ICECT 2027 (NUML).

• Network-in-the-loop testbed: PX4 SITL flights in Isaac Sim (primary) and Gazebo (cross-check) feed unmodified ROS 2 nodes across an NS-3 emulated link (150 ms delay, 50 Mb/s, 0–30% loss).
• Onboard YOLOv8s + ByteTrack sends one 72-byte descriptor per tracked target instead of video.
• Across 200 automated runs, semantic delivery stayed within one point of the 1−p bound (1.00 / 0.96 / 0.86 / 0.70 at 0 / 5 / 15 / 30% loss) at a steady 151 ms and about 1 kB/s. Best-effort 1080p JPEG delivered almost nothing from 5% loss, and reliable JPEG rose to 508–614 ms while using 218–1,310× more bandwidth.

• Generated 6,000 auto-labelled aerial frames (cars, people) in Omniverse Replicator with domain randomization (8–120 m altitude, camera pitch, field of view, lighting) and fine-tuned YOLOv8s against a VisDrone baseline.
• Honest null result: adding synthetic data did not beat real-only fine-tuning (mAP@50 0.372 vs 0.371 on VisDrone val). Synthetic-only training reached 0.364 AP@50 on cars but 0.114 on pedestrians, pointing to the domain gap on small objects.

• Open-source app running PX4 SITL in Isaac Sim (Pegasus) with QGroundControl, several drone models, gimbal control and live YOLO detection on the ROS 2 camera stream (13–14 FPS on an RTX 4060). Reused as the front end of the FANET study.

• Fine-tuned RF-DETR Nano via Roboflow on a custom-annotated aerial dataset of ground vehicles and their components: 89.4% mAP@50, 86.8% precision, 89.2% recall on validation.

• SITL platform coupling Isaac Sim with PX4 and offboard ROS 2 Humble nodes (perception, PID visual servoing) over a bidirectional MAVLink bridge with 50 Hz IMU/state telemetry.

• Ground station that ingests 50–60 Hz telemetry over ZeroMQ in background QThread workers, with a ring buffer for 20 Hz UI rendering and a bypass queue for safety-critical messages.

• Deployed quantized YOLOv8 with BoT-SORT / ByteTrack on Jetson AGX Orin for real-time closed-loop subject tracking in field flights.
• Built a pilot-in-the-loop subject-following system on Raspberry Pi 5 + Coral Edge TPU (24–30 FPS) with Kalman-filtered detections and MAVLink velocity control; flight-tested.
• Built ground control software (PyQt6, .NET, React/Electron) for 50+ Hz MAVLink telemetry, offline-map waypoint missions and threaded YOLOv8n detection on the RTSP feed.
• LiDAR mapping toward GPS-degraded navigation: RPLIDAR S2 + ROS 2 occupancy grids in Gazebo/RViz, tested on a small ground robot.