Stompy's Sandbox

This repository has been replaced with KSim

This repository provides a simple way to train and test RL policies on Stompy, a humanoid robot created by K-Scale Labs. The simulation runs on Mujoco and implements the Gym interface.

Installation

Clone the repository:

git clone https://github.com/michael-lutz/stompy-sandbox.git
cd stompy-sandbox

I recommend creating a new conda enviroment to run the code. Run the following commands to create a new conda environment and activate it:

conda create -n stompy python=3.8
conda activate stompy

Run the following commands to install the required packages in requirements.txt

pip install -r requirements.txt
pip install -e

Usage

To train a policy, navigate to the experiments folder and run the following command:

zsh run_ppo_stompy_walk.sh

Adjust hyperparameters in the any of the .sh files in the experiments folder.

Todo

• Implement MJX to improve scalability
• Speed up rendering! (Currently very slow)
• Collect high quality demonstrations of walking and recovering from falls
• Add AMP discriminator & training setup