pylot

Reinforcement Learning for Flight Simulation

Tanay Biradar, Conner Khudaverdyan, Casey Linden, Edison Zhang

Setup

Create a file .env with the following environment variables:

WANDB_API_KEY=your_api_key
WANDB_ENTITY=your_wandb_username
TOTAL_TIMESTEPS=timesteps_to_train_for

• Training: Run docker-compose build to build the image and docker-compose up to train. The model will be saved to runs
  • To train with GPU, run docker-compose with docker-compose-cuda.yml
• Inference: On your local machine, run: python ppo_inference.py --model-path="path-to-cleanrl_model-file"
  • NOTE: FlightGear must be installed for visualization. Ensure that fgfs --version runs properly

Credits

• CleanRL for baseline PPO implementation
• gym_jsbsim for Gym + JSBSim hooks and environment reward function
• jsbsim
• Richter, D.J., & Calix, R.A. (2021). QPlane: An Open-Source Reinforcement Learning Toolkit for Autonomous Fixed Wing Aircraft Simulation. Proceedings of the 12th ACM Multimedia Systems Conference.
• Note: Everything in utils is not our own code but has been modified to fit the Dockerization needs of this repo. The resulting Docker images are thus self-contained.