ASR-D501 YOLO Drone Demo

Real-time object detection controlling drone motors using CPU and NPU inference on the Qualcomm QCS6490.

Built for Japan IT Week 2026 at the Advantech booth to demonstrate the difference between running YOLOv8 on a general-purpose CPU versus the integrated Hexagon NPU.

---

⚠️ SAFETY WARNING

This is a stationary demonstration only. The drone does not fly. Remove all propellers before running any script. The motors spin to demonstrate that AI detection controls physical actuators - they are not configured for flight. Actual flight-capable scripts with proper safety checks, failsafes, and GPS/altitude hold are not included and would need to be developed separately.

---

What It Does

A camera watches the scene. When the YOLOv8 model detects a clock, the drone motors speed up. Move the clock left or right and the drone yaws to follow. Remove the clock and the motors drop back to idle.

Two versions of the same demo:

• CPU version - YOLOv8n runs on the ARM Cortex-A78 cores (1.3 FPS, 600-700ms latency)
• NPU version - YOLOv8n runs on the Hexagon HTP NPU (110.3 FPS, real-time response)

Same model, same camera, same drone. The only difference is where the inference runs.

---

Hardware

Component	Details
Companion Computer	Advantech ASR-D501 - Qualcomm QCS6490 SoC, 8GB LPDDR5, 128GB UFS, Ubuntu 24.04 LTS
Flight Controller	Pixhawk 1 running ArduCopter V4.6.x
Drone Frame	DJI F450 frame kit - assembly instructions at the link
Camera (Option A)	AnkerWork C310 - USB webcam, outputs MJPEG
Camera (Option B)	Intel RealSense D435i - USB depth camera, RGB stream outputs YUYV

---

Repository Structure

├── README.md                          # This file
├── SETUP-GUIDE.md                     # How to run the demo (start here)
├── docs/
│   ├── ROS2-and-MAVROS-install.md     # Installing ROS 2 Jazzy + MAVROS on Ubuntu 24.04
│   ├── NPU-deployment-guide.md        # Adapting the GStreamer pipeline from Yocto to Ubuntu
│   └── RTABMAP-mapping.md             # 3D mapping with Intel RealSense (optional)
├── scripts/
│   ├── ankerCAM_cpu_demo.py           # CPU inference - AnkerWork C310
│   ├── ankerCAM_npu_demo.py           # NPU inference - AnkerWork C310
│   ├── intelCAM_cpu_demo.py           # CPU inference - Intel RealSense D435i
│   ├── intelCAM_npu_demo.py           # NPU inference - Intel RealSense D435i
│   ├── anker_MJPEG_yolo_cam_ai_hub.sh # GStreamer shell script - AnkerWork (MJPEG)
│   └── intel_YUY2_yolo_cam_ai_hub.sh  # GStreamer shell script - Intel RealSense (YUYV)
└── model/
    ├── labels.txt                     # COCO class labels (80 classes)
    └── yolov8_det.tflite              # Quantized YOLOv8n model for NPU (see below)

---

Quick Start

Full step-by-step instructions: SETUP-GUIDE.md

# Terminal 1 - Start MAVROS
ros2 launch mavros apm.launch fcu_url:=/dev/ttyACM0:115200

# Terminal 2 - Run the demo
cd scripts/

# CPU version (AnkerWork camera):
python3 ankerCAM_cpu_demo.py

# NPU version (AnkerWork camera):
python3 ankerCAM_npu_demo.py

Point the camera at a clock. Motors react.

---

Model Quantization

The NPU requires a quantized TFLite model. The original YOLOv8n PyTorch model (.pt) is converted through:

.pt (PyTorch) → ONNX → .tflite (INT8 quantized for Hexagon HTP)

Follow the Advantech quantization guide: https://github.com/ADVANTECH-Corp/EdgeAI_Workflow/blob/main/ai_system/qualcomm/aom-dk2721/linux/object_detection_demo-using-qc_ai_hub.md

Note: The shell scripts in that guide are written for Yocto Linux with a MIPI camera. If you are running Ubuntu with a USB camera (AnkerWork or RealSense), use the adapted shell scripts in scripts/ instead. See NPU Deployment Guide for details on what changed and why.

---

Available Models

Pre-quantized INT8 TFLite models for the Hexagon NPU are included in the model/ directory:

File	Architecture	Size
yolov8n_det.tflite	YOLOv8 nano	3.3 MB
yolov8s_det.tflite	YOLOv8 small	11.0 MB
yolov8m_det.tflite	YOLOv8 medium	25.2 MB
yolov8l_det.tflite	YOLOv8 large	42.4 MB
yolov8x_det.tflite	YOLOv8 extra-large	67.5 MB
yolov11n_det.tflite	YOLOv11 nano	2.8 MB
yolov11s_det.tflite	YOLOv11 small	9.5 MB
yolov11m_det.tflite	YOLOv11 medium	20.3 MB

The default is yolov8n_det.tflite. To use a different model, change the MODEL_FILE variable at the top of any NPU script:

MODEL_FILE = "yolov8s_det.tflite"  # ← change this

Then copy the model to the scripts directory:

cp model/yolov8s_det.tflite ~/ai-hub/EdgeAI_Workflow/ai_system/qualcomm/aom-dk2721/linux/script/

---

Measured Performance

Metric	CPU	NPU
Inference FPS	~1.3	~110.3
Board power draw	~8.5W	~7.6W
Response latency	~600-700ms (noticeable)	~25ms (instant)
Model size	~12 MB (FP32)	~3.3 MB (INT8)

Idle board power: 5.8W. Board maximum rated: 11W.

---

Prerequisites

• Ubuntu 24.04 LTS on the ASR-D501
• ROS 2 Jazzy + MAVROS installed (installation guide)
• Pixhawk flight controller with ArduCopter firmware
• Camera connected via USB

---

Camera Support

Camera	Format	Device	Decode Stage
AnkerWork C310	MJPEG	/dev/video2	jpegparse → jpegdec → videoconvert
Intel RealSense D435i	YUYV	/dev/video6 (RGB)	videoconvert only

Both cameras produce NV12 after conversion. The NPU inference path is identical from that point on.

To find the correct device for the RealSense:

v4l2-ctl --list-devices
v4l2-ctl -d /dev/videoN --list-formats-ext  # look for YUYV

---

Flight Controller Parameters

These must be set in Mission Planner before running the demo:

Parameter	Value	Purpose
RC_OVERRIDE_TIME	3	RC override timeout (seconds)
ARMING_CHECK	0	Disable pre-arm safety checks
FS_THR_ENABLE	0	Disable throttle failsafe
SYSID_MYGCS	1	Must match MAVROS system_id
BRD_SAFETY_MASK	0	Bypass safety switch
MOT_SPIN_ARM	0.15	Motor spin when armed (15%)
MOT_SPIN_MIN	0.15	Minimum motor spin (15%)

---

Author

Szymon Dudek - Internship at Advantech Co., Ltd., Tokyo, Japan (2025-2026)

Supervisor: Jack Tsao