RLGym-Learn

A flexible framework for efficiently using RLGym v2 to train models.

Features

• Full support for all generics of the RLGym v2 API
• Full support for all functionality of RLGym v2 across multiple environments
• Fast parallelization of environments using Rust and shared memory
• Support for metrics gathering from environments
• Detailed checkpointing system
• File-based configuration
• Provided optimized PPO implementation
• Allows multiple learning algorithms to provide actions for agents within an environment
• Multi-platform (Windows, Linux)

Installation

1. install RLGym via pip install rlgym. If you're here for Rocket League, you can use pip install rlgym[rl-sim] instead to get the RLGym API as well as the Rocket League / Sim submodules.
2. If you would like to use a GPU install PyTorch with CUDA
3. Install this project via pip install git+https://github.com/JPK314/rlgym-learn (coming to PyPI soon)

Usage

See the RLGym website for complete documentation and demonstration of functionality. For now, you can take a look at quick_start_guide.py and speed_test.py to get a sense of what's going on.

Credits

This project was built using Matthew Allen's wonderful RLGym-PPO as a starting point. Although this project has grown to share almost no code with its predecessor, I couldn't have done this without his support in talking through the design of abstractions and without RLGym-PPO to reference. A couple files in this project remain quite similar or even identical to their counterparts in RLGym-ppo - these include:

• experience_buffer.py
• all of standard_impl/util
• actor / critic implementations in standard_impl/ppo

Disclaimer

This framework is designed to be usable in every situation you might use the RLGym API in. However, there are a couple assumptions on the usage of RLGym which are baked into the functionality of this framework. These are pretty niche, but are listed below just in case:

1. The AgentID hash must fit into a signed 64 bit integer.
2. The obs space type and action space type should not change after the associated configuration objects' associated get_x_type functions have been called, and they should be the same across all agents and all envs.