Drone Obstacle Avoidance with AirSim and DDPG

Authors

• Nishant Pandey
• Suriya Suresh

Overview

This project implements a drone obstacle avoidance system using AirSim and the Deep Deterministic Policy Gradient (DDPG) algorithm. The goal is to train a drone to navigate through an environment while avoiding obstacles in real-time.

Packages Used

• stable-baselines3 v1.7.0 (pip install stable-baselines3[extra]==1.7.0)
• airsim v1.8.1
• gym 0.21.0
• Packages as required by the airsim package. A requirements.txt has been attached in case of versions mismatches or the conda env fails to be imported.

Features

• Utilizes the AirSim simulator for realistic drone flight dynamics and sensor data.
• Implements the DDPG algorithm for training the drone to avoid obstacles.
• Provides a user-friendly interface for visualizing the obstacle avoidance behavior.

Contents

There are three folders, lidar, depth , lidar+depth which have train.py, eval.py and drone_env.py. The instructions to run all have been documented below. Other folders and files are :

• readme.md
• env.yml
• requirements.txt
• setting.json

Installation

1. Unzip the folder if you have not done so

2. Install the required packages from the env.yml file using conda

   conda env create -f env.yml

3. Download and setup the AirSim simulator by following the instructions in the AirSim documentation. For this project we used the biniaries provided by airsim under releases , V1.8.1 , AirsimNH.zip.Extract the zip and run the airsimnh.exe for the simulator to work. Also paste the provided config file under Documents/Airsim/settings.json

Usage

1. Launch the AirSim simulator.

2. Navigate to Lidar/Lidar+depth folder, run the main script to start the drone obstacle avoidance system:

   python train.py

3. The drone will start training. Once training is over, it will be saved under /models of the root dir. Logs are saved under /tmp of the root dir which can be viewed using

tensorboard --logdir=/tmp/name_of_folder

Evaluation

1. Launch the AirSim simulator.

2. Modify the model_path in the eval.py file to the path of your model. Run the main script to start the drone obstacle avoidance system:

   python  eval.py

3. Evaluation starts and you can see the output in the airsim window. After evaluation is done, we can see the metrics which show up in a matplotlib window and also is saved in the root directory.

Results

The results qualitative and qunatitative can be seen in the report.pdf uploaded in the root of this repo.

• This shows the results of Lidar-Only implmenation. It can be seen that that the drone in many episodes narrowly escapes the thin wooden pole and tree.

• This shows the results of Lidar+Depth Camera Implementation. This shows showcases many instances when the drone actively tries to change its path to avoid obstacles.

Bugs

• If the code does not connect to the simulator, please change the port in settings.json under ApiServerPort as well as on the code , in the files drone_env_ddpg.py and ddpg_lidar.py in line number 19 for both eval and the normal files.

Acknowledgements

• AirSim - Open-source simulator for drones and cars developed by Microsoft.

• DDPG - Deep Deterministic Policy Gradient algorithm for continuous control tasks.

• UAV_Navigation_DRL_AirSim for opensourcing his code as it helped us understand a lot and we have used parts of his code in our codebase.

• DDPG-AirSim-Drone-Obstacle-Avoidance for helping us understand DDPG and how to control a drone in airsim using RL.

• Reinforcement-learning