DRLoco - DeepMimic meets MuJoCo for Locomotion

Note

DRLoco = DeepMimic(MuJoCo, StableBaselines3)

This repository for you, when:

• you have a bipedal robot and need a controller to generate stable human-like walking

• you have a MuJoCo model of a bipedal robot and you want it to learn to walk like a human

• you have any kind of a legged robot and a set of reference trajectories (e.g. motion capturing) and you want a robust controller to let your robot imitate the reference motion.

• you want to try out the DeepMimic Approach with minimal effort

• you need high quality mocap data for Imitation Learning

How to read this documentation?

There might be something worth explaining.

TOC

• Repository Overview
  • File Structure
  • Folder drloco
  • Folder mujoco
  • Other Folders
• Installation
  • Installing Additional Features
• Mimic Environment
  • Purpose
  • Main Functions
• Reference Trajectories
  • Overview
  • Details
  • How to use your own reference trajectories?
• Motion Capturing Data
• Walker Environments
  • Straight Walker
  • 2-segmented upper body and 3d hip walker
  • How to use your own environment?
• Trained Models
• Configuration and Hyperparameters
  • Training Configuration
  • Hyperparameters and PPO modifications
• Monitoring Training
  • Callback
  • Monitor Environment Wrapper
  • Weights & Biases (WB) and Tensorboard (TB)
  • How to log an additional custom metric to TB and WB?

Scripts

• Training an agent
  • Training Script
  • Training Procedure
• Running an agent
  • WHEN do we save agents?
  • WHAT info do we save for each agent?
  • WHERE do we save trained agents?
• Evaluating the performance of an agent

Background

• Deep Mimic Framework
  • Reference State Initialization (RSI)

Indices and tables

• Index

• Module Index

• Search Page