Generative Synthesis of Kinematic Mechanisms (NeurIPS 2025 Creative AI Track)

📄 Paper 📦 Dataset 🌐 Project Website

Overview

This project implements Variational Autoencoders (VAE) for generating kinematic mechanisms from curves and vice versa. The system supports both CNN-based and Vision Transformer (ViT) based architectures for learning the relationship between mechanism curves and their corresponding linkage structures.

Project Structure

GenMech/
├── model/                    # Core model components
│   ├── dataset.py           # Dataset and data loading utilities
│   ├── loss.py             # Loss functions for VAE training
│   ├── VAE_CNN.py          # CNN-based VAE implementation
│   ├── VAE_VIT.py          # Vision Transformer VAE implementation
│   ├── VIT_decoder_timm.py # ViT decoder using timm
│   └── validation.py       # Validation utilities
├── test/                    # Evaluation and testing scripts
│   ├── eval_test.py        # Model evaluation and testing
│   └── eval_plot.py        # Results visualization and plotting
├── mechanism/              # Dataset generation pipeline
├── train.py               # Main training script
└── requirements.txt       # Python dependencies

Setup Instructions

1. Clone the Repository

git clone https://github.com/jl6017/GenMech.git
cd GenMech

2. Environment Setup

# Create a new conda environment
conda create -n genmech python=3.11

# Activate the environment
conda activate genmech

# Install PyTorch with CUDA support (adjust for your system)
conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia

# Install remaining dependencies
pip install -r requirements.txt

3. Data Preparation

Dataset follows this structure:

dataset/
├── tri_2/
│   └── images/
│       ├── curve/        # Curve images (.png)
│       └── mechanism/    # Mechanism images (.png)
├── complex_t4/
│   └── [subdirectories with same structure]
└── test_0807/
    └── [test data with same structure]

Usage

Training

# Run training with default parameters
python train.py

# The script will automatically:
# - Load data from ../dataset/ (adjust path in train.py if needed)
# - Train either ViT or CNN model (configured in script)
# - Save results and checkpoints to results/ directory

Evaluation

# Run model evaluation
python test/eval_test.py

# Generate learning curve plots
python test/eval_plot.py

Key Training Parameters

• Model Type: Switch between ViT and CNN in train.py
• Dataset: Configure DATASET and DATAPATH variables
• Hyperparameters: Adjust beta, gamma, learning_rate etc.

Recommend Env

• Python 3.11
• PyTorch 2.0+
• CUDA 12.4

Citation

If you use this code, please cite:

@article{lin2025creative,
  title={Creative synthesis of kinematic mechanisms},
  author={Lin, Jiong and Ning, Jialong and Goldfeder, Judah and Lipson, Hod},
  journal={arXiv preprint arXiv:2510.03308},
  year={2025}
}