Game Theoretic Forecasting and Planning

Website | Paper

This is an official implementation for the IROS 2023 paper:
A Game Theoretic Framework for Joint Forecasting and Planning
Kushal Kedia, Prithwish Dan, Sanjiban Choudhury

We implement a game-theoretic framework for joint forecasting and planning on the CrowdNav environment, and compare with the industry standard of Maximum-Likelihood Estimation (MLE) based forecasting and planning.

Setup

Setup environments following the SETUP.md

Training

Train MLE-Forecaster and Nominal Planner

cd crowd_nav
python train_mle_forecaster.py
python train_nom_planner.py

Finetune the above models using the game-theoretic framework.

python train_forecaster_planner_game.py

Evaluation

The following code compares the costs and collision rates of our approach with MLE.

cd crowd_nav
python evaluate.py

Results

Safe Planning with Adversarial Forecasts

Collision with MLE Forecasting and Planning

Work in Progress

We will soon release the implementation of our algorithm on the ETH-UCY benchmark.

Acknowledgement

This repository borrows code from Social-NCE.

Citation

@inproceedings{kedia2023game,
  title={A Game-Theoretic Framework for Joint Forecasting and Planning},
  author={Kedia, Kushal and Dan, Prithwish and Choudhury, Sanjiban},
  booktitle={IROS},
  year={2023}
}