Flow RL is a high-performance reinforcement learning library, combining modern deep RL algorithms with flow and diffusion models for advanced policy parameterization, planning ability or dynamics modeling. It features:
- State-of-the-Art Algorithms and Efficiency: We provide JAX implementations of SOTA algorithms, such FQL, BDPO, DAC and etc;
- Flexible Flow Architectures: We provide built-in support various types of flow and diffusion models, such as CNFs and DDPM;
- Comprehensive Evaluations: We test the algorithms on commonly adopted benchmark and provide the results.
Currently FlowRL is hosted on PyPI and therefore can be installed via pip install flowrl. However, we recommend to clone and install the library using the following commands:
git clone https://github.com/typoverflow/flow-rl.git
cd flow-rl
pip install -e .The entry files are presented in examples/. Please refer to the scripts in scripts/ for how to execute the algorithms.
Offline RL:
| Algorithm | Location | WandB Report |
|---|---|---|
| IQL | flowrl/agent/iql.py |
[Performance] [Full Log] |
| IVR | flowrl/agent/ivr.py |
[Performance] [Full Log] |
| FQL | flowrl/agent/fql/fql.py |
[Performance] [Full Log] |
| DAC | flowrl/agent/dac.py |
[Performance] [Full Log] |
| BDPO | flowrl/agent/bdpo/bdpo.py |
[Performance] [Full Log] |
If you use Flow RL in your research, please cite:
@software{flow_rl,
author = {Chen-Xiao Gao and Mingjun Cao},
title = {Flow RL: Flow-based Reinforcement Learning Algorithms},
year = 2025,
version = {v0.0.1},
url = {https://github.com/typoverflow/flow-rl}
}Inspired by foundational work from