Introduction to MuJoCo

MuJoCo is a free and open-source physics engine designed to facilitate research and development in robotics, biomechanics, graphics, and animation, providing fast and accurate simulation capabilities.

• MuJoCo: Multi-Joint dynamics with Contact

• Rigid-body simulation

• Essential simulator components:

  • Solving equations of motion (e.g., articulated body algorithms)
  • Contact solver (key factor for sim-to-real gaps)
  • User-friendly visualizer

Tutorial Topics

1. Introduction to MuJoCo and mujoco_parser
2. Forward Kinematics
3. Forward Dynamics
4. Inverse Kinematics
5. Inverse Dynamics
6. Reinforcement Learning with Soft Actor-Critic for Snapbot

Environment Setup

Follow the steps below to configure your environment

1. Create and Activate Conda Environment

Create a new conda environment and activate it:

conda create -n snapbot-env python=3.10
conda activate snapbot-env

2. Install Dependencies

Install all required packages from the provided requirements.txt file:

pip install -r requirements.txt

---

Project: Snapbot Olympics

Overview

Optimize robotic motion using reinforcement learning. Train and evaluate control policies in simulation for various competition tasks.

Events

• Standing Long Jump: Forward distance achieved by robot after jumping
• High Jump: Vertical height achieved by robot
• Running Sideways: Time to reach designated goal coordinate (faster is better)

Simulation

• All tasks conducted in MuJoCo

Provided Resources

• Snapbot robot model (XML)
• SAC algorithm implementation (notebook/06_sac_snapbot_train.ipynb, notebook/06_sac_snapbot_eval.ipynb, pakage/gym/snapbot_env.py, pakage/rl/sac.py)
• (Reward functions must be designed by students)

Tasks

• Configure neural network dimensions based on robot DoF and action dimensions
• Design task-specific reward functions
• Train and evaluate policies
• Analyze reward function impact on performance

Submission

• One-page A4 result report (must include reward definitions and model hyperparameters)
• Video of trained Snapbot performance

Aim to design the best policy and win the Snapbot Olympics!

---

Contact

• Prof. Sungjoon Choi: sungjoon dash choi at korea dot ac dot kr
• TA Taemoon Jeong: taemoon dash jeong at korea dot ac dot kr