Latent Importance Sampling

This repository contains code and experiments for the paper:

Kruse, L. A., Tzikas, A., Delecki, H., Arief, M., & Kochenderfer, M. J., Enhanced Importance Sampling through Latent Space Exploration in Normalizing Flows, AAAI 2025. [arXiv]

Dependencies

See environment.yaml for required packages, or use this to create a Conda environment with all dependencies:

conda env create -f environment.yaml

This repository was tested with Python 3.11 and PyTorch 2.1. The racecar and f16 experiments require additional installations (discussed in the following section) if you want to retrain a flow model or generate a dataset of Monte Carlo simulations.

Data

The experiments expect a {simulator}-flow.csv and {simulator}-mcs.csv file for each simulator (robot, racecar, f16). Generated datasets can be downloaded at this link, and the .csv files should be placed into a folder called data/.

New datasets (for additional flow training or Monte Carlo evaluations) can be created using the files in the datagen/ folder. For the nonholonomic robot experiments, run the following command:

python3 datagen/robot-data.py

Generating data for the racecar experiments requires a Julia install and the MPOPIS (Model Predictive Optimized Path Integral Strategies) repository from the following paper:

@inproceedings{Asmar2023},
  title = {Model Predictive Optimized Path Integral Strategies},
  author = {Dylan M. Asmar and Rasalu Senanayake and Shawn Manuel and Mykel J. Kochenderfer},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year = {2023}

Specific installation instructions can be found in the MPOPIS repository.

Generating data for the f16 experiments requires a jax install and the Jax F16 Dynamics repository from the following paper:

@inproceedings{So-RSS-23},
  title = {Solving Stabilize-Avoid Optimal Control via Epigraph Form and Deep Reinforcement Learning}, 
  author = {Oswin So AND Chuchu Fan}, 
  booktitle = {Proceedings of Robotics: Science and Systems}, 
  year = {2023}, 
} 

Specific installation instructions can be found in the Jax F16 repository.

Flow Training

Pre-trained flows are provided in the flows/ folder. However, if you wish to train a flow from scratch, you can run the following command, replacing the simulator argument with the simulator of choice (robot, racecar, f16):

python3 train.py --simulator 'robot'

Flow hyperparameters are stored in yaml files in the configs/ folder.

Latent Importance Sampling Experiments

To reproduce the experimental results, run the experiments.sh file. Importance sampling hyperparameters are stored in yaml files in the configs/ folder. To view the logged results (in either tabulate format or in LaTeX format), run

python3 utils/results_table.py

How to Cite

If you find this code useful in your research, please cite the following publication:

@inproceedings{kruse2025enhanced,
  title={Enhanced Importance Sampling through Latent Space Exploration in Normalizing Flows},
  author={Kruse, Liam Anthony and Tzikas, Alexandros and Delecki, Harrison and Arief, Mansur and Kochenderfer, Mykel J},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={17},
  pages={17983--17989},
  year={2025}
}