Autonomous VTOL UAV

View Source
PythonC++ROSCVRaspberry Pi

A custom-built Vertical Take-Off and Landing (VTOL) Unmanned Aerial Vehicle designed for fully autonomous missions. This project integrates embedded systems, real-time telemetry, and computer vision to create a drone capable of self-correction and intelligent pathfinding.

🛰️ System Architecture

The system is split into two main components: the Onboard Flight Systems and the Ground Control Station.

Onboard Hardware

  • Compute: Raspberry Pi 4 (8GB) handling high-level logic and video encoding.
  • Flight Controller: Pixhawk running ArduPilot for low-level stabilization.
  • Communication: LoRa Transceiver for long-range telemetry and RC control.
  • Sensors: GPS, Barometer, IMU, and a dedicated Pi Camera for visual input.

Ground Control Station (GCS)

The "brain" of the autonomy stack resides on a high-performance laptop acting as the GCS.

  1. Telemetry Link: Receives mavlink packets via the transceiver.
  2. Video Stream: Decodes low-latency video feed transmitted from the RPi.
  3. Inference Engine: Runs object detection and path planning algorithms on the GPU.
  4. Command Relay: Sends corrected flight vectors back to the UAV in real-time.

🧠 The Autonomy Pipeline

The core challenge was closing the loop between visual perception and flight control over a wireless link.

graph LR
    A[Raspberry Pi Camera] -->|Encoded Video| B[Transceiver]
    B -->|Wireless Link| C[Laptop GCS]
    C -->|Inference Model| D[Path Planner]
    D -->|MAVLink Commands| B
    B -->|Control Signal| E[Pixhawk FC]

1. Visual Navigation

The Raspberry Pi streams 720p video to the GCS. Using OpenCV and a custom YOLO model, the laptop identifies landing zones and obstacles. The decision to offload processing to the ground station allowed for much heavier models than the Pi could run onboard.

2. Control Loop

Once a target is identified, the laptop calculates the necessary roll, pitch, and yaw adjustments. These are converted into MAVLink commands and sent back to the drone. The latency was optimized to be under 100ms to ensure stable flight characteristics.

🚧 Challenges

Bandwidth vs. Latency

Streaming video and telemetry simultaneously over a limited bandwidth radio link required aggressive compression. I implemented a custom UDP protocol to prioritize the latest frame, dropping dropped packets immediately rather than re-requesting them.

Failsafe Mechanisms

Reliance on a ground station introduces a critical point of failure: connection loss. To mitigate this, the onboard Raspberry Pi runs a lightweight "Keep-Alive" watchdog. If the link is severed for more than 500ms, the drone automatically switches to a "Return to Launch" (RTL) mode using its onboard GPS.

🔮 Future Improvements

  • Onboard Edge AI: Moving inference to a Coral Edge TPU to remove the dependency on the ground station.
  • Swarm Logic: Enabling communication between multiple UAVs for coordinated formation flight.