[DCL] Drone Calibration Lab

The repository for the UT Dallas Drone Calibration Lab.

The DCL is built on top of the PyBullet engine, and serves as a testing environment for cheaply and quickly prototyping combat strategies for drone-on-drone warfare. The DCL also offers benchmarks for AI drone controllers in various tasks.

---

Installation

Requirements

• We recommend the MiniForge CLI for virtual environment and package management. MiniForge is a much more efficient version of Anaconda, and dramatically improves speed and overall performance. MiniForge is also pre-configured to work with the popular conda-forge channel. You can download the Miniforge installer for your device here. Scroll down to the "Download" section, and choose a version that's right for you.
  • If you still wish to use anaconda, simply replace the "mamba" command with the "conda" command and everything should still work the same, only slower.
  • If you use anaconda, you might need to specify for conda to use the conda-forge channel for some of our required libraries. This can be done by adding the text -c conda-forge onto the end of your installation command.
• Git and your IDE of choice is recommended for cloning/modifying the project onto your local computer. You can download Git here.
• IMPORTANT NOTE: The AI model trained by the simulator requires an NVIDIA GPU for use.

Create an Environment

• Open MiniForge Prompt and create your virtual environment. Our code was built using Python 3.8, so we recommend this version.

  mamba create --name dcl "python=3.8"
  mamba activate dcl
  

Install Required Libraries

• Library	Command	Stable Version
  python		3.8.18
  mamba*	Comes pre-installed in MiniForge	22.3.1
  pytorch	mamba install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia	2.2.1
  opencv	mamba install opencv	4.0.1
  matplotlib	mamba install matplotlib	3.7.2
  pybullet	mamba install pybullet	3.25
  pynput	mamba install pynput	1.7.6
  einops	mamba install einops	0.7.0
  gstreamer	mamba install gstreamer	1.18.5

---

Quick Start

How to Run

• Edit your desired changes to CONFIG.py in VSCode/IDE of choice.
  • There, you can change episode and hitpoly parameters, as well as any other configurations.
• Run Main.py with your desired specifications and appropriate action flag.

  python Main.py --action generate_data
  

Usage

python Main.py --action [OPTIONS]

python Main.py -a [OPTIONS]

Action Flag Options

• help  Output a usage message and exit.
• build_video  Creates an .mp4 video visualizing the change in hitpoly for the drone based on its position and direction after completion of data generation and training.

Data Generation

• generate_data  Generates initial data for training. Attributes such as episode_count and episode_length may be modified.
• stitch_data  Stitches together tensors generated from generate_data.

Hitpoly Training

• train_hitpoly  Trains the hitpoly model in batches. Hitpoly parameters such as epochs, learning rate, and batch size may be modified.
• query_hitpoly  Loads the model output from raw pixel array values to be rendered as color images.
• render_hitpoly  Generates the hitpoly simulator starting from episode 1. Hold the enter key to progress each frame.

---

FAQ

It says a GPU is required. Can I run this without a GPU?

No. If you were to attempt to run the simulator anyway, you would receive a runtime error. A dedicated GPU is necessary for the high loads of data and processing.

Do I need Mamba? I've never seen this library before.

If this is your first time installing Mamba, you may choose to not use it and instead stick with conda instead. This will be considerably slower, however, as conda is quite old and has not aged well.

Error: "Torch not compiled with CUDA enabled."

You may have a CPU-only Pytorch version previously installed. Try installing pytorch in a clean MiniForge environment via mamba, and see if that resolves the issue.

Error: "Initializing libiomp5md.dll, but found libiomp5md.dll already initialized."

This is an old anaconda error which can be fixed by deleting the libiomp5md.dll file located in anaconda3/envs/dcl/Library/bin/, as is described in this stack overflow post.

---