piSTAR Lab
An open-source platform built to make deep reinforcement learning experimentation accessible and fun.
piSTAR Lab is a modular platform for training and comparing reinforcement learning agents. You can run everything from a web UI or do the same things from Python.
- A web UI for setting up agents, environments and experiments
- An extension system for adding new agents, environments and task types
- Single-player and multi-agent environments
- Experiment tracking, with episode recordings
- Distributed processing using Ray under the hood
It shipped with Landia, a small game world I built for agents to play in. The project isn’t actively developed anymore, but the code and docs are still up.