/RLGymPPO_CPP

A lightning-fast C++ implementation of RLGym-PPO

Primary LanguageC++

RLGymPPO_CPP

A lightning-fast C++ implementation and extension of RLGym-PPO, as well as rlgym-sim

Speed

Results will vary depending on hardware, but it should be substantially faster for everyone.

On my computer (Intel i5-11400 and GTX 3060 Ti), this repo is about 5x faster than Python RLGym-PPO on default settings. Collection has the most substantial benefit, and I can reach upwards of 70ksps on my computer in C++, vs 10k in Python.

Features

This implementation adds several features that RLGym-PPO/rlgym-sim doesn't have (mostly because it does not fit in Aech's scope for RLGym-PPO):

  • Different multithreaded collection model that doesn't require constant cross-thread communication
  • Can run far more environments (hundreds to thousands) simultaneously using far less memory per environment
  • Actual multithreading used instead of separate processes and shared memory
  • Fully-configurable skill tracking system using ELO
  • Full RocketSim CarState/BallState access in GameState (e.g. player.carState.isFlipping)
  • RocketSim Arena access in GameState
  • Built-in zero-sum rewards with adjustable opponent scale
  • Built-in padded obs builder with slot shuffling
  • Support for more advanced state setters via RocketSim Arena access
  • Added possibility for rewards to override their behavior across all players
  • Support for collection during learn
  • Support for auto-casted learn
  • Better multithreading of learn for CPU-only
  • Built-in gradient noise measurement system
  • Added callbacks for steps and iterations
  • Uses RocketSim Vec class with various quality-of-life functions like .Length(), .Dist(), etc.

Accuracy to Python RLGym-PPO

According to several different learning tests, RLGymPPO_CPP and RLGym-PPO have no differences in learning.

Installation

  • Clone this repository recursively: git clone https://github.com/ZealanL/RLGymPPO_CPP --recurse
  • If you have an NVIDIA GPU, install CUDA 11.8: https://developer.nvidia.com/cuda-11-8-0-download-archive
  • Download libtorch for CUDA 11.8 (or for CPU if you don't have an NVIDIA GPU): https://pytorch.org/get-started/locally/
  • Put the libtorch folder inside RLGymPPO_CPP/RLGymPPO_CPP
  • Open the main RLGymPPO_CPP folder as a CMake project (if you're on Windows, I recommend Visual Studio with the C++ Desktop package)
  • Change the build type to RelWithDebInfo (Debug build type is very slow and not really supported) (don't worry you can still debug it)
  • Make sure you have a global Python installation with wandb installed (unless you have turned off metrics)
  • Build it
  • Add your collision_meshes folder to wherever the executable is running

Transferring models between C++ and Python

You can do this using the script tools/checkpoint_converter.py

I've confirmed that this script works perfectly, however you will need to make sure the obs builder and action parser match perfectly in Python

Dependencies