ReaL LSD This repository was developed as a semester project at the Autonomous Systems Lab ETH Zurich by Nasib Naimi, and expanded by Nicola Loi and Valentin Ibars as a project for the course "Perception and Learning for Robotics". Extending on work done in gym-unrealcv, the repository was built to investigate the feasibility of deep reinforcement learning for solving the problem of landiing MAVs in emergency situations. A custom scene and MAV model were built in Unreal Engine which served as the environment and agent respectively. The agent was trained using Proximal Policy Optimization (PPO) and truncated Generalized Advantage Estimation (GAE).
The real_lsd package extends and inherits classes from the Gym-UnrealCV package and as such depends on it. Install gym-unrealcv and setup the package using
git clone https://github.com/zfw1226/gym-unrealcv.git
cd gym-unrealcv
pip install -e .
and install OpenCV. If you are using anaconda or miniconda, see the section on installation in the README of gym-unrealcv.
Installing the real_lsd setting up the package can be done similarly
git clone https://github.com/nanaimi/real-lsd.git
cd real-lsd
pip install -e .
Once the real-lsd repository is installed, copy the game binary used as the environment for training and testing the agent from real-lsd/real_lsd/envs/UnrealEnvs/ into 'gym-unrealcv/gym_unrealcv/env/UnrealEnv/' using
cp -R <yourpath>/real-lsd/real_lsd/envs/UnrealEnvs/<Binary> <yourpath>/gym-unrealcv/gym_unrealcv/env/UnrealEnv/
Following the instructions found on OpenAI's Gym Registration, custom environment can be registered by using register() at load time. This is done in the __init__.py file of real_lsd and can be seen here.
To test UnrealCV in a packaged game built on Unreal Engine, download any game binaries built with the UnrealCV Plugin. Examples can be found in the model zoo here.
After downloading the game binary, start the game and press the key ` twice to open the unrealCV console. Test different commands for interacting with the game in the console. Available requests can be found in the UnrealCV Wiki.
Alternatively, you can test the UnrealCV python client by launching the examples/test.ipynb notebook.
The implementation of the PPO agent can be found under agents/ppo_agent, including the python executable that trains the agent and runs it through multiple episodes for testing. Run the following to train a new agent with PPO and test it in 200 episodes and specify the path to which the data collected throughout training should be stored.
python agents/ppo_agent/PPO_lander.py
Test a trained agent in the custom scene built for showing a proof a concept.
TODO
- add model model parameters
- add demo notebook
TODO
CUDA = 11.0