DiffFR: Differentiable SPH-based Fluid-Rigid Coupling for Rigid Body Control

1. Introduction

This is the code base for our SIGGRAPH Asia 2023 paper. We develop our method based on the open-source library SPlisHSPlasH and we sincerely thank the effort of its authors.

2. Installation

We have built and tested the code on Ubuntu22.04.

2.1 Prerequisite

Required infrastructures can be installed via:

sudo apt install git cmake xorg-dev freeglut3-dev build-essential libomp-dev

Required python packages can be directly installed by pip install -r requirements.txt.

2.2 How to build

First shallow clone this repo with git:

git clone --depth=5 git@github.com:zhehaoli1999/DiffFR.git 

Then we recommend to create a new conda environment (such as "diffFR") and activate it:

conda create --name diffFR python=3.7
conda activate diffFR

Then find build.sh under the root dir of the project, and use commands like bash build.sh to execuate this shell (please ensure your shell has already execuated conda init). It will use cmake to configure the project and then compile it with your compiler.

3. Experiments in paper

You can find all raw data records and the scripts for figure plotting of our experiments in the experiments folder.

The simulation scene settings are described as json files, and the initial setting of velocities and positions of fluid and rigid particles are stored as bgeo and binary files.

To run each task, you need to input a scene json file, as well as an optional binary file of the initial settings of the fluid environment (without specification, all fluid particles will have the positions described in the json file with zero velocities. )

3.1. Rigid body trajectory optimization

Data

| Task | Scene json file | state file | taskType | design variables | notes | |---------------------------|--------------------------------|-----------------------------|------------------|------------------|-------------------------------------------------| | water rafting | diff-water-rafting-bunny.json | water_rafting/state_130.bin | 'water-rafting' | vx, vz omega | load-fluid-pos-and-vel | | high diving | diff-high-diving-duck.json | high_diving/state_98.bin | 'high-diving' | omega | load-fluid-pos | | stone skipping | diff-stone-skipping.json | stone_skipping/state_18.bin | 'stone-skipping' | v, omega | load-fluid-pos | | bottle flipping (stage 1) | diff-bottle-model.json | bottle_flip/state_54.bin | 'bottle-flip' | v, omega | load-fluid-pos | | bottle flipping (stage 2) | diff-bottle-model-collide.json | bottle_flip/state_54.bin | 'bottle-flip' | v, omega | load-fluid-pos | | on-water billiards | billiards-on-water-2balls.json | billiards/state_17.bin | - | v, omega | use diff-rigid-contact-multi.py, load-fluid-pos |

Commands to run

First, cd ./experiments/rigid_body_trajectory_optimization/python.

For gradient-based optimization (ours), use gradient-based-optimize.py. For gradient-free optimization (CMA-ES & 1+1 ES), use opt-ng.py.

• bottle flipping

python gradient-based-optimize.py --scene=../scene/diff-bottle-model.json --state=../state/bottle_flip/state_54.bin  --load-fluid-pos --taskType=bottle-flip

• stone skipping

python gradient-based-optimize.py --scene=../scene/diff-stone-skipping.json --state=../state/stone_skipping/state_18.bin  --load-fluid-pos --taskType=stone-skipping

• water rafting

python gradient-based-optimize.py --scene=../scene/diff-water-rafting-bunny.json --state=../state/water_rafting/state_130  --load-fluid-pos-and-vel --taskType=water-rafting 

• high diving

python gradient-based-optimize.py --scene=../scene/diff-high-diving-duck.json --state=../state/high_diving/state_98.bin   --load-fluid-pos --taskType=high-diving

• on-water billiards

python diff-rigid-contact-multi.py --scene=../scene/billiards-on-water-2balls.json --state=../state/billiards/state_17.bin  --load-fluid-pos

4.2. Self-supervised bottle flipping control policy

Commands to run

First, cd ./experiments/bottle_flip_robot/python

python bottle-flip-robot.py --scene=../scene/diff-bottle-model-360degree.json --state=../state/state_1.365999.bin --load-fluid-pos

4.3. On-water inverted pendulum robot controller

Commands to run

First, cd ./experiments/on_water_inverted_pendulum/python

Train:

python cartpole-diff-controller.py --scene=../scene/cartpole-diff-controller.json --state=../state/state_0.879992.bin --load-fluid-pos

Test:

python cartpole-diff-controller.py --scene=../scene/cartpole-diff-controller.json --state=../state/state_0.879992.bin --load-fluid-pos --test=../saved_models/success8s-2023-01-14-23-07-05-cartpole-epoch6.pth

4.4. Other comparison & ablation experiments

Please see the content of ./experiments/others for code and raw data.

5. Postprocess & Rendering

For postprocess we use Houdini to reconstruct the fluid surface mesh from particles, and use Blender Cycles for rendering.

6. Contact & citation

If you have any question, please feel free to open an issue or email zhehaoli@mail.ustc.edu.cn.

To cite our paper, here is the bibtex:

@article{10.1145/3618318,
author = {Li, Zhehao and Xu, Qingyu and Ye, Xiaohan and Ren, Bo and Liu, Ligang},
title = {DiffFR: Differentiable SPH-Based Fluid-Rigid Coupling for Rigid Body Control},
year = {2023},
issue_date = {December 2023},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
volume = {42},
number = {6},
issn = {0730-0301},
url = {https://doi.org/10.1145/3618318},
doi = {10.1145/3618318},
journal = {ACM Trans. Graph.},
articleno = {179},
numpages = {17}
}

7. License

MIT License