Learning Time-Critical Responses for Interactive Character Control

Abstract

This code implements the paper Learning Time-Critical Responses for Interactive Character Control. This system implements teacher-student framework to learn time-critically responsive policies, which guarantee the time-to-completion between user inputs and their associated responses regardless of the size and composition of the motion databases. This code is written in java and Python, based on Tensorflow2.

Publications

Kyungho Lee, Sehee Min, Sunmin Lee, and Jehee Lee. 2021. Learning Time-Critical Responses for Interactive Character Control. ACM Trans. Graph. 40, 4, 147. (SIGGRAPH 2021)

Project page: http://mrl.snu.ac.kr/research/ProjectAgile/Agile.html

Paper: http://mrl.snu.ac.kr/research/ProjectAgile/AGILE_2021_SIGGRAPH_author.pdf

Youtube: https://www.youtube.com/watch?v=rQKuvxg5ZHc

How to install

This code is implemented with Java and Python, and was developed using Eclipse on Windows. A Windows 64-bit environment is required to run the code.

Requirements

Install JDK 1.8

Java SE Development Kit 8 Downloads

Install Eclipse

Install Eclipse IDE for Java Developers

Install Python 3.6

https://www.python.org/downloads/release/python-368/

Install pydev to Eclipse

https://www.pydev.org/download.html

Install cuda and cudnn 10.0

CUDA Toolkit 10.0 Archive

NVIDIA cuDNN

Install Visual C++ Redistributable for VS2012

Laplacian Motion Editing(PmQmJNI.dll) is implemented in C++, and VS2012 is required to run it.

Visual C++ Redistributable for Visual Studio 2012 Update 4

Install JEP(Java Embedded Python)

Java Embedded Python

This library requires a part of the Visual Studio installation. I don't know exactly which ones are needed, but I'm guessing .net framework 3.5, VC++ 2015.3 v14.00(v140). Installing Visual Studio 2017 or later may be helpful.

Install Tensoflow 1.14.0

pip install tensorflow-gpu==1.14.0

Install this repository

We recommend downloading through Git in Eclipse environment.

1. Open Git Perspective in Elcipse
2. Paste repository url and clone repository ( 'https://git.ncsoft.net/scm/private_khlee/private-khlee-test.git' )
3. Select all projects in Working Tree
4. Right click and select Import Projects, and Import existing Eclipse projects.

Or you can just download the repository as Zip file and extract it, and import it using File->Import->General->Existing Projects into Workspace in Eclipse.

Install third party library

This code uses Interactive Character Animation by Learning Multi-Objective Control for learning the student policy.

Download required third pary library files(ThirdPartyDlls.zip) and extract it to mrl.motion.critical folder.

Dataset

The entire data used in the paper cannot be published due to copyright issues. This repository contains only minimal motion dataset for algorithm validation. SNU Motion Database was used for martial arts movements, CMU Motion Database was used for locomotion.

How to run

Eclipse

All of the instructions below are assumed to be executed based on Eclipse. Executable java files are grouped in package mrl.motion.critical.run of project mrl.motion.critical.

• You can directly open source file with Ctrl+Shift+R
• You can run the currently open source file with Ctrl+F11.
• You can configure program arguments in Run->Run Configurations menu.

Pre-trained student policy

You can see the pre-trained network by running RuntimeMartialArtsControlModule.java. Pre-trained network file is located at mrl.python.neural\train\martial_arts_sp_da

• 1, 2 : walk, run
• 3,4,5,6 : martial arts actions
• q,w,e,r,t : control critical response time

How to train

1. Data Annotation & Configuration
   • You can check motion data list and annotation information by executing MAnnotationRun.java.
2. Model Configuration
   • Action list, critical response time of each action, user input model and error metric is defined at MartialArtsConfig.java
3. Preprocessing
   • You can precompute data table for pruning by executing DP_Preprocessing.java
   • The data file will be located at mrl.motion.critical\output\dp_cache
4. Training teacher policy
   • You can train teacher policy by executing LearningTeacherPolicy.java
   • The result will be located at mrl.motion.critical\train_rl
5. Training data for student policy
   • You can generate training data for student policy by executing StudentPolicyDataGeneration.java
   • The result will be located at mrl.python.neural\train
6. Training student policy
   • You can train student policy by executing mrl.python.neural\train_rl.py
   • You need to set program arguments in Run->Run Configurations menu.
     • arguments format : <new/load>
     • ex) martial_arts_sp new 0.0001
7. Running student policy
   • You can see the trained student policy by running RuntimeMartialArtsControlModule.java.
   • This class will be load student policy located at mrl.python.neural\train.