Multi-modal UAV for Natural Disaster Rescue

A hybrid locomotion UAV system capable of both flying and driving for natural disaster rescue scenarios, based on RL-pybullets-cf. (Flight functionality is implemented, and we additionally implemented driving functionality on top of it.)

Overview

This project implements a system that automatically selects between driving and flying modes to maximize drone energy efficiency in u-shaped environment (A terrain with flat ground between two cliffs). The drone is equipped with both propellers and wheels, enabling efficient hybrid locomotion.

Key Features

• Hybrid Locomotion:

  • Wheeled driving for energy-efficient ground movement
  • Quadcopter flight
  • Automatic mode selection based on environment

• Intelligent Control:

  • Regression-based mode selection (Flight or drive). It selects a way to reach the target based on cliffs's height and distance between them.
  • Trajectory following with RL

• Environment Analysis:

  • U-shaped map structure analysis
  • Energy-efficient mode prediction (94.4% accuracy)

System Architecture

1. Hardware Design

   • 4 propellers for flight control
   • 4 wheels for ground locomotion
   • Integrated control system for mode switching

2. Control System

   • Wheel-Propeller coordination
   • Trajectory following
   • Mode transition management
   • Energy optimization selection between flight and drive

3. Decision Making

   • Environment analysis
   • Mode selection by regression model

Setup

Tested on ArchLinux, Ubuntu and macOS. Note that Eigen must be installed on the system.

On Ubuntu:

bash sudo apt-get install libeigen3-dev

On macOS:

bash brew install eigen

It is strongly recommended to use a python virtual environment such as conda or pyvenv.

Installation Steps

1. Clone repository (recursively)

bash git clone https://github.com/rolsj/CS470_trajectory_following.git git submodule update --init --recursive

1. Setup Python environment (tested with Python 3.10.13)

bash pyenv install 3.10.13 pyenv local 3.10.13 python3 -m venv ./venv source ./venv/bin/activate pip3 install --upgrade pip

1. Install dependencies and build

bash pip3 install -e . # Ubuntu users might need: sudo apt install build-essential

Usage

Scripts for training, testing and visualization are provided.

Training

bash cd runnables ./train_rl.sh

Testing

bash cd runnables ./test_rl.sh ./test_pid.sh

Visualization

bash cd runnables ./vis_rl.sh

Run experiments

```bash ./runnables/exper_rl.sh [mode] [h1] [h2] [l]

mode = drive or flight // h1, h2, l are float. Example 1: ./runnables/exper_rl.sh drive 1.5 1.5 8 A terrain is created with two cliffs, each 1.5 meters high, and a 7-meter gap between them. Currently, h1 and h2 must be the same. ```

Implementation Details

The system is implemented using PyBullet physics engine with custom URDF models for the hybrid drone:

1. Hybrid Locomotion Control

Ground Movement

• Velocity-based differential wheel control
• Dynamic speed adjustment based on future waypoints
• Automatic direction control using cross-product for optimal rotation
• Logarithmic speed scaling for smooth acceleration/deceleration
• Integrated ground contact detection for mode switching

Flight Control

• 4-motor RPM control system
• Attitude-based control
• Smooth transition between ground and flight modes

2. Path Planning & Navigation

Trajectory Generation

• Parabolic trajectory generation for both modes
• Flight mode: Curved path
• Ground mode: Curved paths from cliff to ground. (Different parabolic path is generated based on the cliff.) A direct path connecting takeoff and landing.
• Dynamic waypoint generation based on terrain

Mode Selection

• Energy consumption prediction for each mode
• Automatic mode switching based on:
  • Distance to target
  • cliff's heights

3. Environment Generation

Map Generation

• Parametric U-shaped terrain generation
• Configurable cliff heights (h1, h2)
• Adjustable cliff distance (l)
• Dynamic cliff URDF model generation

Energy Model

• Mode-specific energy consumption calculation
• Ground mode: Energy is calculated by multiplying the torque acting on the wheel by its angular velocity.
• Flight mode: It is calculated using the formula for energy used by the propeller.
• Real-time energy monitoring and logging

Results

1. Mode Selection Efficiency

   • Drive mode optimal for longer distances
   • Flight mode preferred for short distances
   • 94.4% accuracy in mode prediction

2. Energy Optimization

   • Significant energy savings through mode switching
   • Optimal path planning for minimal energy consumption

Team

• Doheun Kim
• Minsol Park
• Doun Lee
• Seongjun Lee

KAIST