Research Internship - Crafting Npcs & Bots Behaviors With Llm - Vlm - F - M - Nb H/F Ubisoft

Recent advancements in Large Language Models (LLM) and Visual Language Models (VLM) are revolutionizing many fields, including video games. Multiple video-game companies are already studying how to leverage LLMs for game-changing innovations, such as enabling players to cooperate in natural language with NPCs (non-player characters). Within academia, multiple research labs are studying how to adapt LLMs/VLMs to new use cases, such as interacting in complex environments through reinforcement learning or evolutionary strategies. These breakthroughs open up vast opportunities for enhancing player immersion through more intelligent and responsive NPCs.

This internship aims to explore the integration of LLMs and VLMs with reinforcement learning (RL) or evolutionary strategies to enhance the intelligence and behavior of NPCs and bots in video games. By combining the creative potential of LLMs with the interactive feedback loops of RL, we aim to build better NPCs and reduce iteration time for game developers.

The internship will take place within Ubisoft's La Forge, where cutting-edge tools and resources are available for research, including a set of realistic 3D games specifically designed for research purposes, datasets, and state-of-the-art learning algorithms deployed on GPU clusters. This internship will benefit from close collaboration with the Inria Flowers team, led by Pierre-Yves Oudeyer.

Objectives:

- Investigate the current uses of LLMs/VLMs in gaming, focusing on how they enhance NPC decision-making.
- Study recent research on using reinforcement learning to improve the performance of LLM/VLMs, with a particular focus on grounding these models in interactive environments such as video games.
- Throughout discussion with production teams, identify a potential real-world use case on which to calibrate prototyping and research.
- Experiment with methods to optimize NPC learning in simulated or live gaming environments, providing a balance between realism and computational efficiency.