PeDesCar

This intends to generate datasets related to impact collision using Mujoco Physics Engine, OpenAI, RL for human motion prediction.

Accepted Paper – WSCG 2025 Explore our latest research and reproducible results: Paper PDF

Sample Data is available in the sample_data folder.

For the full dataset and collaboration opportunities, reach out via: QualityMinds LinkedIn

Installation

tested on python 3.8.10

Copy the mujoco license in the bin folder.

First be sure you have the linux package required to run the environment:

sudo apt-get install libglew-dev
pip install wheel setuptools pip --upgrade

Clone this repo:

git clone https://github.com/QualityMinds/PeDesCar.git

Install package:

pip3 install -r requirements.txt
pip3 install -e pedestrian-impact/ # to install the environments

Install imitation_lib:

git clone https://github.com/robfiras/ls-iq.git
cd ls-iq
python3 -m pip install -e .
cd ../

Install loco-mujoco and get dataset:

git clone --branch v0.1.0 https://github.com/robfiras/loco-mujoco.git
cd loco-mujoco
python3 -m pip install .
loco-mujoco-download
cd ../

Copy pretrained agents file:

cp -R /mnt/nfs-nasmonster/ai-research/notebooks/kiattention/impact_collision_dataset/agents pedestrian-impact/pedestrian_impact/envs/

Run (main files)

Run code for dataset generation.

python dataset_generation.py "logs/" 5 1

Run code for dataset visualization

python dataset_visualizer.py "path-experiment-experiment" 0 --save-numpy

Run code for dataset analysis

python dataset_analysis.py "path-experiment-experiment" 0

Run code for figure plots (ongoing and not validated) path should be changed in the code.

python figures_for_paper.py

TODO

We made a plan composed of the following milestones:

• Car modelling.
• Replace OpenAI Human model by MyoHuman model.
• Run Env with MyoHuman Human model.
• Train a Imitation algorithm on MyoHuman to learn motions.
• Run Env with MyoHuman Human model using the learned motions.