UAV Autonomous Landing and Takeoff at Portable Vertiports

Overview

This project explores the use of Deep Reinforcement Learning (DRL) to enable autonomous landing and takeoff of unmanned aerial vehicles (UAVs) at portable vertiports. The study investigates a novel approach for multi-UAV coordination in dynamic environments, focusing on minimizing flight times and avoiding collisions during simultaneous landing and takeoff operations.

Table of Contents

1. Introduction
2. System Architecture
3. Simulation Environment
4. Deep Reinforcement Learning Model
5. Training and Testing
6. Results
7. How to use the code

Introduction

Unmanned aerial vehicles (UAVs) have become pivotal across numerous applications such as package delivery, remote sensing, and emergency communications. This project introduces a portable vertiport system, which can be deployed flexibly to reduce unnecessary energy consumption and enable continuous operations. The goal is to autonomously manage the simultaneous takeoff and landing of multiple UAVs in a constrained environment, using advanced reinforcement learning techniques.

System Architecture

Vertiport Model

The vertiport consists of:

• A central spinal frame supporting multiple cylindrical landing pads.
• Foldable, encapsulated landing pads to shield UAVs from adverse weather.

UAV Model

Each UAV is represented as a sphere with predefined aerodynamic properties

Simulation Environment

The environment is modeled in a 3D coordinate system measuring 5 × 5 × 5 meters. It includes:

• Multiple UAVs initiating takeoff and landing sequences from predefined locations.
• Non-Cooperative Entities (NCEs) such as birds or amateur drones, which introduce dynamic obstacles.

Deep Reinforcement Learning Model

The core of this project is a DRL model designed to operate in a Markov Decision Process (MDP) framework. Key features include:

• Observations: State information from UAVs and NCEs.
• Actions: Acceleration values optimized for collision avoidance and efficient maneuvering.
• Rewards: Designed to incentivize successful landings and penalize collisions or out-of-bound movements.

Training and Testing

Training Setup

• The model was trained using an NVIDIA GeForce GTX 3060 GPU and an Intel i7 processor.
• UAVs were trained to avoid dynamic obstacles and reach their destinations efficiently.

Experiment Parameters

• A grid-based positioning strategy was used for initial placements and destinations.
• Mobile NCEs moved at a constant speed of 0.1 m/s with random destinations, mimicking real-world scenarios.

Results

Section to be completed based on experimental outcomes.

How to use the code

To be finished