Minimal PPO Implementation

pip install minppo

We provide a minimal implementation of training a robotic agent to do simple tasks, using the Brax framework.

Installation

Brax is a versitile physics engine for training agents in reinforcement learning that can work on your local computer or on a cluster. The easiest way to install the minppo package is to use pip:

pip install minppo

Command Line Interface

There are some default configurations in the minppo/configs directory which you can use to get started:

minppo train stompy_pro

Alternatively, you can specify the path to the config file directly:

minppo train configs/stompy_pro.yaml

Additionally, you can use dot notation to override specific parameters:

minppo train configs/stompy_pro.yaml 'rl.num_env_steps=1000'

You can visualize the environment with:

minppo env stompy_pro

Finally, you can run inference on a trained model with:

minppo infer stompy_pro 'inference.model_path=path/to/model.pkl'