Deep RL for Fluid-Rigid Control

Deep-reinforcement-learning Controlled Fluid-Rigid Simulation

A python-based ball-shooting game of fluid-rigid simulation and autonomous control via Deep Reinforcement Learning.

Project Link on GitHub

Table of Contents

• Features
• Installation
• Usage
• Demo

Features

• MLS-MPM method for fluid simulation
• Newton-Eulerian system for rigid body movement
• Lagrangian sampling for fluid-rigid coupling
• Robustness and Vram optimization for GPU parallelization
• Soft Actor-Critic deep reinforcement learning framework for tube control
• Meta Learning for problem generalization
• CNN Autoencoder pretraining for correlated field properties

Installation

Requirements

• python 3.10

• jittor 1.3.6

• taichi 1.4

• gym 0.26.2

• pytorch 2.1.0

Platform

Good Nvidia GPU (cuda) to run simulation, rendering and learning

Usage

How to play with shooting tube

python play.py

• A(←) & D(→): control the tube to move leftwards and rightwards
• W(↑) & S(↓): increase or decrease the ejecting speed
• right & left mouse click: control the tube to rotate clockwise and counter-clockwise
• R: reset the ball and tube

Quick Start

Train

python train.py ./configs/train.json

Eval

python eval.py ./configs/eval.json

-provided models in model_trial1

Demo

-Effects of the AutoEncoder:

—— by Zhiwei Zhao, Dec. 2023