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 expectation is that the agent navigates through these openings without colliding with blocks. This project currently runs only on Windows since Unreal environments were packaged for Windows.

• Stable Baselines 3
• OpenAI - Gym
• Microsoft AirSim
• Unreal Engine 4

The training environment has 9 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.

• State is in the form of a RGB image taken by the front camera of the agent.
• Image shape: 50 x 50 x 3

• There are 9 discrete actions.

#️⃣ 1. Clone the repository

git clone https://github.com/bilalkabas/PPO-based-Autonomous-Navigation-for-Quadcopters

#️⃣ 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 ppo_drone python==3.8

#️⃣ 3. Install required libraries

Inside the main directory of the repo

conda activate ppo_drone
pip install -r requirements.txt

#️⃣ 4. (Optional) Install Pytorch for GPU

You must have a CUDA supported NVIDIA GPU.

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": 20,
    "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
            }
        ]
    }
  }

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

So, inside the repository

python main.py

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

So, inside the repository

python policy_run.py

The trained model in saved_policy folder was trained for 280k steps.

The test environment has different textures and hole positions than that of the training environment. For 100 episodes, the trained model is able to travel 17.5 m on average and passes through 4 holes on average without any collision. The agent can pass through at most 9 holes in test environment without any collision.

• Bilal Kabas

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