This repository contains our work on creating automatic safety software for autonomous vehicles at MakeHarvard 2020. Big thanks to https://github.com/abhinavsagar/Reinforcement-Learning-Tutorial for a great implementation of continuous space Deep-Q learning. We use this to train our RL model. To get started, run:
conda install pytorch -c pytorch
pip install -r requirements.txt
The navigate to rc_car_env and run:
pip install -e .
This makes our custon environment and physics engine. The source code for that is in rc_car_env/rc_car_env/envs/.
To train the RL model, go to Reinforcement-Learning-Tutorial/ddpg walker/DDPG.ipynb and run that notebook. Move the resulting .pth files into the model directory.
To see how the model performs, run python simulation.py simple 1000. This shows (in a very toy case) how the model safety checks the usual moving algorithm and cuts it off before it runs into the obstacle. simple represents the toy case, and regular is the usual paramater. The other two arguments are steps and episodes respectively. A new episode lays out a new random environment.
Lastly, to interact with the physical RC car, connect to the car's onboard wifi server and run python application.py. Navigate to 127.0.0.1:5000 and use the (rather crummy) interface to drive the car. The model should kick in and prevent you from ramming into a wall or other kind of obstacle.