Skip to content

Deep Reinforcement Learning based autonomous navigation in realistic simulation environments.

License

Notifications You must be signed in to change notification settings

bilalkabas/DRL-Nav

Repository files navigation

DRL-Nav: Autonomous UAV Navigation via Deep Reinforcement Learning Using PPO

license

ℹ️ This work is an extension of my previous work: "PPO-based Autonomous Navigation for Quadcopters". In this work, actions are continuous and the agent is trained using different input types. Tests are performed both randomly and sequentially.

⚠ This document is best viewed in light theme.

This repository contains an implementation of Proximal Policy Optimization (PPO) for autonomous navigation in a corridor environment with a quadcopter. There are blocks having circular opening for the drone to go through for each 4 meters. The goal is that the agent navigates through these openings without colliding with blocks. The train and test environments were created using Unreal Engine and Microsoft AirSim. This project currently runs only on Windows since Unreal environments were packaged for Windows.

Contents

Overview

The training environment has 15 sections with different textures and hole positions. The agent starts at these sections randomly. The starting point of the agent is also random within a specific region in the yz-plane.

Inputs

There are three models trained using depth, single RGB, and stacked gray images, respectively. Their sizes are as follows

  • Depth map: 50 x 50 x 1
  • Single RGB image: 50 x 50 x 3
  • Depth image: 50 x 150 x 1

Actions

There are two actions:

equation

Neural Network

In this work, a five layer neural network is used.


Results

The test environment has different textures and hole positions than that of the training environment.

Random tests

Average success rate (%) in random tests for each models trained using three different input types:

Input type Train Test
Depth %100 %99.5
Single RGB %98.5 %92
Stacked gray %98.5 %96
Random actions %11 %22.5

Sequential Tests

Trajectories of the agent trained using depth and stacked gray images (medium and difficult test environments are indicated by blue and red, respectively):

Environment setup to run the codes

#️⃣ 1. Clone the repository

git clone https://github.com/bilalkabas/DRL-Nav

#️⃣ 2. From Anaconda command prompt, create a new conda environment

I recommend you to use Anaconda or Miniconda to create a virtual environment.

conda create -n drl_nav python==3.8

#️⃣ 3. Install required libraries

Inside the main directory of the repo

conda activate drl_nav
pip install -r requirements.txt

#️⃣ 4. (Optional) Install Pytorch for GPU

You must have a CUDA supported NVIDIA GPU.

Details for installation

For this project, I used CUDA 11.0 and the following conda installation command to install Pytorch:

conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch

#️⃣ 4. Edit settings.json

Content of the settings.json should be as below:

The setting.json file is located at Documents\AirSim folder.

{
    "SettingsVersion": 1.2,
    "LocalHostIp": "127.0.0.1",
    "SimMode": "Multirotor",
    "ClockSpeed": 40,
    "ViewMode": "SpringArmChase",
    "Vehicles": {
        "drone0": {
            "VehicleType": "SimpleFlight",
            "X": 0.0,
            "Y": 0.0,
            "Z": 0.0,
            "Yaw": 0.0
        }
    },
    "CameraDefaults": {
        "CaptureSettings": [
            {
                "ImageType": 0,
                "Width": 50,
                "Height": 50,
                "FOV_Degrees": 120
            },
            {
                "ImageType": 2,
                "Width": 50,
                "Height": 50,
                "FOV_Degrees": 120
            }
        ]
    }
}

How to run the training?

Make sure you followed the instructions above to setup the environment.

#️⃣ 1. Download the training environment

Go to the releases and download TrainEnv.zip. After downloading completed, extract it.

#️⃣ 2. You can change the training mode to produce model outputs for different input types

In the main project directory, go to config.yml. Here you can change the training mode to depth, single_rgb, or multi_rgb.

#️⃣ 3. Now, you can open up environment's executable file and start the training

So, inside the repository

python train.py

How to run the pretrained model?

Make sure you followed the instructions above to setup the environment. To speed up the training, the simulation runs at 20x speed. You may consider to change the "ClockSpeed" parameter in settings.json to 1.

#️⃣ 1. Download the test environment

Go to the releases and download TestEnv.zip. After downloading completed, extract it.

#️⃣ 2. Change the test mode

In config.yml, you can change the test mode to depth, single_rgb, or multi_rgb. This should match the input type that the model was trained with. In the same config file, you can change the test type to sequential or random.

#️⃣ 3. Now, you can open up environment's executable file and run the trained model

So, inside the repository

python inference.py

Citation

@INPROCEEDINGS{9864769,  
author={Kabas, Bilal},  
booktitle={2022 30th Signal Processing and Communications Applications Conference (SIU)},   
title={Autonomous UAV Navigation via Deep Reinforcement Learning Using PPO},   
year={2022},  
pages={1-4},  
doi={10.1109/SIU55565.2022.9864769}}

Author

License

This project is licensed under the GNU Affero General Public License.

About

Deep Reinforcement Learning based autonomous navigation in realistic simulation environments.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages