UAV_Navigation_DRL_AirSim

This is a new repo used for training UAV navigation (local path planning) policy using DRL methods.

ChangeLog

• 2022-03-11
  • Add wandb support
• 2022-03-10
  • Remove gym_airsim_multirotor submodule
  • Add gym_env as envrionment, include MultirotorSimple, Multirotor and FixedwingSimple dynamics
  • Add train with plot
  • Add SimpleAvoid UE4 environment

Requirements

• Python 3.8
• AirSim v1.6.0
• pytorch 1.10.1 with gpu
• gym-0.21.0
• Pyqt5 5.15.6
• keyboard 0.13.5

Submodules

• Stable-baselines3 v1.4.0

Install CUDA and PyTorch (Win10)

• Download CUDA11.6
• pip3 install torch==1.10.1+cu113 torchvision==0.11.2+cu113 torchaudio===0.10.1+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
• You can use scripts/test/torch_gpu_cpu_test.py to test your PyTorch and CUDA

Usage

1. Clone this repo and the submodules

   1. git clone https://github.com/heleidsn/UAV_Navigation_DRL_AirSim.git --recursive

2. Install gym_env

   1. cd gym_env
   2. pip install -e .

3. Install customized stable-baselines3

   1. cd stable-baselines3
   2. pip install -e .

4. Download a AirSim environment, such as Blocks from Here and run it

5. Start training

   1. cd UAV_Navigation_DRL_AirSim
   2. python scripts/start_train_with_plot.py

6. Evaluation

   1. cd UAV_Navigation_DRL_AirSim
   2. python scripts/start_evaluate_with_plot.py

Configs

This repo using config file to control training conditions.

Now we provide 3 training envrionment and 3 dynamics.

env_name

• SimpleAvoid

  • This is a custom UE4 environment used for simple obstacle avoidance test. You can download it from google drive.

  

• City_400_400

  • A custom UE4 environment used for fixedwing obstacle avoidance test. You can also get it from google drive

  

• Random obstacles

  • Some envs with random obstacles. Contributed by Chris-cch. You can download here.

  

• Other Airsim build in envrionment (AirSimNH and CityEnviron):

  

dynamic_name

• SimpleMultirotor
• Multirotor
• SimpleFixedwing

GUI for training and evaluation

Wandb support

Wandb is a central dashboard to keep track of your hyperparameters, system metrics. You can find examples here.

Note: If you use wandb, please run python as administators.

Results

Training result using TD3 with no_cnn policy

Benchmark

2D depth navigation Benchmark for 3 different algorithms and 5 different policies in SimpleAvoid environment:

Settings

Note:

To speed up image collection, you can set ViewMode to NoDisplay.

For multirotor with simple_flight controller, please set SimMode to Multirotor. You can also set ClockSpeed over than 1 to speed up simulation (Only useful in Multirotor mode).

Also, it's better to put your environment files in your SSD rather than HDD.

{
  "SeeDocsAt": "https://github.com/Microsoft/AirSim/blob/master/docs/settings.md",
  "SettingsVersion": 1.2,
  "SimMode": "ComputerVision",
  "ViewMode": "NoDisplay",
  "ClockSpeed": 1,
  "SubWindows": [
    {"WindowID": 0, "CameraID": 0, "ImageType": 0, "Visible": true},
    {"WindowID": 1, "CameraID": 0, "ImageType": 3, "Visible": false},
    {"WindowID": 2, "CameraID": 0, "ImageType": 3, "Visible": true}
    ],
  "CameraDefaults": {
    "CaptureSettings": [
      {
        "ImageType": 3,
        "Width": 100,
        "Height": 80,
        "FOV_Degrees": 90,
        "AutoExposureSpeed": 100,
        "AutoExposureBias": 0,
        "AutoExposureMaxBrightness": 0.64,
        "AutoExposureMinBrightness": 0.03,
        "MotionBlurAmount": 0,
        "TargetGamma": 1.0,
        "ProjectionMode": "",
        "OrthoWidth": 5.12
      },
      {
        "ImageType": 0,
        "Width": 256,
        "Height": 144,
        "FOV_Degrees": 90,
        "AutoExposureSpeed": 100,
        "AutoExposureBias": 0,
        "AutoExposureMaxBrightness": 0.64,
        "AutoExposureMinBrightness": 0.03,
        "MotionBlurAmount": 0,
        "TargetGamma": 1.0,
        "ProjectionMode": "",
        "OrthoWidth": 5.12
      }
    ]
  }
}