Interaction-Aware Sampling-Based MPC with Learned Local Goal Predictions

This repository contains the training and testing scripts for the prediction model used in the "Interaction-Aware Sampling-Based MPC with Learned Local Goal Predictions" paper.

Data

To download the data and trained models (prediction model and pre-trained autoencoder), run

./download_data.sh

Setup

This repository requires python <= 3.7 and tensorflow == 1.15.x. The instructions were tested on Ubuntu 16.

./install.sh
source social_vdgnn/bin/activate

Training

The training script for the model is src/train_roboat.py. In order to train a new model on the data, specify a model number (--exp_num) and the training parameters in /src/train.sh. An example is provided in the file. Then in /src, run

./train.sh

Testing

In a similar fashion, specify training parameters for the trained model in /src/test.sh (examples provided) and run

./test.sh

Model parameters will be automatically saved in /trained_models

If you find this code useful, please consider citing:

@inproceedings{jansma2023,
  author={Jansma, Walter and Trevisan, Elia and Serra-Gómez, {\'A}lvaro and Alonso-Mora, Javier},
  booktitle={2023 International Symposium on Multi-Robot and Multi-Agent Systems (MRS)}, 
  title={Interaction-Aware Sampling-Based MPC with Learned Local Goal Predictions},
  year={2023},
}

This code and the model are base on "Social-VRNN: One-Shot Multi-modal Trajectory Prediction for Interacting Pedestrians" by de Brito et al. You can find their code here and their paper here. If relevant, please consider citing them too.