Reinforcement Learning in Bevy Rust

This repository is a project to explore rust and reinforcement learning at the same time

Setup

todo: recall where I got libtorch and add that to setup instructions and generally setup project to run (will figure this out when I setup on desktop)

Running

Environment Variables

Set the following environment variables in terminal from the projects root directory before running

export LIBTORCH_BYPASS_VERSION_CHECK=1
export LIBTORCH="$(pwd)/libtorch"
export DYLD_LIBRARY_PATH="$(pwd)/libtorch/lib" # mac
export LD_LIBRARY_PATH="$(pwd)/libtorch/lib" # windows/linux

Startup Command

cargo run --bin main

Flags:

to use flags run the startup command cargo run --bin main -- --<flag-1> --<flag-2>

• --headless : runs game with no window if passed
• --ai-control : specify weather a human or ai is playing

AI Model

to build the ai model architecture, run python model_arc.py from the directory src/modeling

Devlog

Plan

Create ball sorting game/simulation, and train pytorch model in the environment

I would like to define the model architecture in python using pytorch then load into rust to train using RL. See here.

---

Here is getting the basic bevy scene set up with a court full of balls bouncing around

---

I've got the core functionality of the game set up here with a player ball meant to sort the colored balls into their respective quadrants

---

Now we have the (untrained) AI model controling the player ball over several episodes

---

(still untrained) Its efficient to run train on batched trajectories, this allows for higher learning rates