Learning Sampling Distribution and Safety Filter for Autonomous Driving with VQ-VAE and Differentiable Optimization

This repository contains the source code to reproduce the experiments in our IROS 2024 paper where we compare CVAE based model to VQVAE based model for trajectory planning.

Refer to this repo for the CVAE baseline used in our comparison.

Getting Started

1. Clone this repository:

git clone https://github.com/cisimon7/VQOptMain.git
cd VQOptMain

2. Create a conda environment and install the dependencies:

conda create -n venv python=3.8
conda activate venv
pip install -r requirements.txt

3. Download Trained Models and save to directory named weights

4. Download the training dataset and extract the files to the dataset directory, if you want to retrain.

Reproducing our main experimental results

Running Experiment

Using the main.yaml file, you can configure the environment and optimizer parameters or you can run with the command line args as below:

python3 inferenceJax.py 'env={seed: 42, speed_limit: 15, num_eps: 50, render: False, plot: False}' 'optimizer={maxiter: 100, v_des: 20}' 'batch_size=1000'

Comparison using the simulator

VQVAE

CVAE