GAME Data & Artificial Intelligence

Research projects

r&d lab

AXE 1 : GAME Data & Artificial Intelligence

AI is historically linked to video games, being present at all stages of creation, from design through development to production.  In this area, teacher-researchers are particularly interested in three themes. 

1 - AI for game content :

  • Procedural generation of game content (levels, terrain, scenery, characters, etc.)
  • Automated game testing
  • Generative AI engines

2 - AI for Gameplay :

  • Algorithms for real-time decision makingand action planning of Non-Player Characters (NPCs)
  • Algorithms for perceptions
  • Generative‘AI engines for Gameplay and new game concepts

3 - Game data :

  • In-game data collection
  • Data analysis
  • Modelling player behaviour
  • etc

This axis is a member of the Games & AI Working Group of the CNRS and naturally linked to the Game Programming course during which students propose, develop and investigate original topics applied to concrete case studies. 

Keywords: Decision making, action planning, game data, procedural content generation, generative AI. 

projects

projects & publications

1 - Procedural generation of Rush Hour puzzle games

Context 

 

Procedural generation of puzzle games allows for more varied and diverse levels, providing a better gaming experience. However, deep learning-based generation is very difficult to implement for this type of game due to the inherent complexity of their design rules.

 

Topic

 

In this work, we propose a model composed of three main modules: a solver and two classifiers. The components of Rush Hour are generated randomly, and the solver aims to evaluate the playability of the generated levels by classifying them as solvable/unsolvable and assessing the difficulty of the solvable levels.

 

Two neural networks are used to improve generation. One is trained to differentiate solvable levels from unsolvable levels, and the other is trained to classify the difficulty of the levels. The robustness of the proposed approach is examined on 6×6 and 7×7 grids. The results obtained showed the effectiveness of our modules in generating interesting and varied levels with four degrees of difficulty for the Rush Hour puzzle game, where the classifier was able to quickly identify unsolvable levels.

 

Our approach could easily be adapted and extended to procedurally generate other types of games, such as platform games or RogueLike games.


Below are three examples of Rush Hour levels generated by our model (the blue blocks represent the cars that need to be moved to unblock the red car, and the black blocks are fixed).

puzzle game_rush hour

Keywords: procedural generation, deep learning, puzzle games

Partner: This project was carried out in partnership with the LAMSADE laboratory at Paris Dauphine University.  

Years 2021/2022 (project completed)

Publication: Computers and Games: International Conference, November 22–24, 2022, Springer-Verlag, Berlin  https://doi.org/10.1007/978-3-031-34017-8_15 

Context

 

NPC behavior is a crucial element in rogue-like games: the more unpredictable the behavior, the greater the player’s immersion. Action planning algorithms can generate such behavior. However, the quality of the game and the immersion do not depend solely on this behavior.

Several other elements can have a major impact on immersion, such as levels, terrain, music, lighting, rendering, and textures, etc. In this project, we investigate the use of Game AI Planning for procedural online level generation.

 

Topic

 

We are studying the ability of an action planner based on Goal Oriented Action Planning (GOAP) to achieve a dual objective:

  1. controlling NPC behavior;
  2. automatically generating dungeon levels.
 

To evaluate this approach, we developed a custom 2D platform game called Khaldun in Unity, inspired by Dead Cells. The game can be downloaded here. The results showed GOAP’s ability to automatically generate solvable dungeons. Each new session generates a unique level, providing an enjoyable and entertaining immersion (see below).

 

In addition, our algorithm produces unpredictable enemy actions, thus avoiding redundant behaviors as in the case of conventional decision-making systems (state machines or behavior trees).

Despite simultaneous generation, planning execution time remains negligible, allowing for a smooth, enjoyable, and entertaining experience.

A room whose main structures (Tiles) are sketched out by the Game Designer.

The room corresponding to the one sketched is obtained with the decorations.

Example of an entire level generated randomly with our GOAP approach by connecting different rooms to each other.

Keywords: platform game, action planning, procedural level generation, GOAP

Partner: Saint-Cyr Coëtquidan Military Academy

Years 2020/2021 (project completed)

Publication : 2021 IEEE Conference on Games (CoG), Copenhagen, Denmark, 2021, pp. 1-2, doi: 10.1109/CoG52621.2021.9619062   

Context

 

In recent years, generative artificial intelligence has challenged the video game industry. Although AI-generated content cannot replace the final content developed and refined by humans in the video game production pipeline, it will likely accelerate the pre-production and prototyping phases, thereby reducing production costs.

In this context, we believe that several targeted tasks will benefit from generative AI. Thus, we believe that integrating local AI tools into video game engines is an avenue worth actively pursuing. One possible task is the generation of decision-making AI for non-player characters (NPCs).

 

prompt

Topic

 

In this study, we propose using local, lightweight LLM language models to generate behavior trees from natural language text. Game developers can describe the desired behavior tree via a prompt window, and our tool will automatically generate it with the option to iterate on it with new prompts.

This tool was developed for Unreal Engine. The LLM is loaded locally on the PC’s graphics card and must therefore be relatively lightweight (compared to current models running on large servers and consuming a lot of energy). Developers can provide the LLM with the decision tree tasks they have available in addition to the basic tasks available in Unreal.

This allows the LLM to adapt to the specific situation to generate the behavior tree first in XML and then directly as an Unreal asset.

project setting
unreal engine

Keywords: LLM, BT, decision-making AI, generative AI, Unreal Engine

Partner: Institute for Information Processing (TNT), Leibniz University Hannover

Years 2024/2026 (project in progress) 

Context

 

Fighting games provide challenging simulation environments for many decision-making problems that model the behavior of non-player characters (NPCs).

The most commonly used methods are based on scripts, state machines, or behavior trees and have been integrated into game engines such as Unreal Engine or Unity.

However, these methods can suffer from redundant and predictable actions, reducing player immersion. The action planning approach overcomes these limitations. However, it is known to be costly in terms of time and memory resources. This therefore poses a challenge for real-time brawler-type games.

 

Topic

 

In this study, we present a hybrid planning approach that combines the advantages of two methods:

  • goal-oriented action planning (GOAP)
  • hierarchical task networks (HTN).

 

The proposed approach is evaluated through a dynamic 3D arena “combat” game offering a rich universe where the player controls a character who controls a fluid material to attack enemies (spiders or crabs). The game can be downloaded here.

The performance of our architecture was analyzed in two scenarios:

  1. the player has high health and is fighting a single spider (Figure 1);
  2. the player is damaged and must fight 3-4 spiders (Figure 2).

 

The execution time for the first configuration is negligible, while it is only 0.01 ms for the second situation despite its unfavorable scenario (multiple NPCs). Therefore, the proposed hybrid approach is effective and does not affect the fluidity of the game even though it runs on the PS4 console.

kahiora_figure 1
kahiora_figure 2

Figure 1: situation 1

Figure 2: situation 2

Keywords: 3D Fighting Game, Action Planning, GOAP, HTN

Partner: Saint-Cyr Coëtquidan Military Academy

Years 2020/2021 (project completed)

Publication : 2021 IEEE Conference on Games (CoG), Copenhagen, Denmark, 2021, pp. 1-2, doi: 10.1109/CoG52621.2021.9619075.

Context

 

Generative AI such as ChatGPT, Mid-Journey, and DALL-E continue to develop in various technological fields. Although the video game industry is somewhat reluctant to use them at the moment, these tools could ultimately be involved in the production of a video game, from design and development to the final gameplay. The main question we are seeking to answer here is: can these AIs be integrated into video game design tools (video game engines) and thus offer players a new gameplay experience?

 

Topic 

 

The goal of this project is to offer players an interactive experience by incorporating a large language model (LLM) such as ChatGPT into the Unity3D game engine. This integration allows players to interact with 3D content intuitively using natural language, where they write and send their prompts to ChaGPT. The generative AI then transmits these instructions to Unity3D to execute them. This allows players to perform actions such as changing the colors, shapes, or functions of 3D elements in the game.

The ChatGPT-Unity3D latency has been optimized to ensure a smooth gaming experience. A video showing a sequence from the game is available below.

This demonstrator has shown the feasibility of considering ChatGPT as a new gaming experience, with intuitive and simple gameplay. Other generative AI tools could be explored in the future to develop new game and gameplay concepts.

Partner: Studio Oh BiBi

Years 2022/2023 (project completed) 

These projects are developed as part of the students’ Game Programming coursework. They are regularly presented at national conferences, mainly at events organised by the CNRS Games & AI working group or others.

  • Apply
  • contact