2-player Ping Pong game with Kuka iiwa R820 arm + gantry

Requirements

• Python 3.12
• macOS (Apple Silicon or Intel) or Ubuntu 20.04+

Installation

1. Create and activate an environment

macOS:

python3.12 -m venv venv
source venv/bin/activate

Ubuntu (conda):

Install Miniconda if not already installed, then:

conda create -n sim2real python=3.12 -y
conda activate sim2real

2. Install MuJoCo

pip install mujoco

MuJoCo 3.x is installed as a Python package — no separate binary installation needed.

3. Install remaining dependencies

pip install stable-baselines3 torch gymnasium scipy numpy

Ubuntu only — install OpenGL/GLFW system libraries required by MuJoCo's renderer:

sudo apt install -y libgl1-mesa-dev libglfw3 libglew-dev

4. (macOS only) Set up mjpython

MuJoCo's passive viewer requires the mjpython launcher on macOS to ensure the window is created on the main thread. It ships with the mujoco package:

# Find mjpython
python -c "import mujoco, os; print(os.path.join(os.path.dirname(mujoco.__file__), 'mjpython'))"

Add the printed directory to your PATH permanently:

# Add to ~/.zshrc
export PATH="$(python -c "import mujoco, os; print(os.path.dirname(mujoco.__file__))")":$PATH

Running

macOS — use mjpython instead of python, otherwise the viewer crashes with NSWindow should only be instantiated on the main thread:

mjpython comp.py

Ubuntu — plain python works:

python comp.py

SB3 deserialization warnings

When loading the PPO model you may see:

UserWarning: Could not deserialize object clip_range / lr_schedule.
Exception: code() argument 13 must be str, not int

These are harmless. They occur because the model was saved on Python ≤3.10 and the code() constructor signature changed in Python 3.12. Policy inference is unaffected.

File structure

lfr-project/
├── comp.py                  # Main competition script
├── mujoco_env_comp.py       # MuJoCo Gymnasium environment
├── model_arch.py            # SimpleModel neural network definition
├── models/
│   ├── model1.pth           # Learned Q-value model for player 1 (This is pretty old, may be wrong, please consider training new one)
│   ├── model2.pth           # Learned Q-value model for player 2 (This is pretty old, may be wrong, please consider training new one)
│   └── model_p.pth          # Nash-p joint strategy model (This is pretty old, may be wrong, please consider training new one)
└── logs/
    └── best_model_tracker_slow/best_model   # PPO policy checkpoint (I just trained this, so should be good to run comp.py wiht random strategy)