Multiplayer Interactive World Models with Representation Autoencoders
Anthony Hu, Václav Volhejn, Adrien Ramanana Rahary, Chris Mulder, Aditya Makkar, Amélie Royer, Manu Orsini, Alyx Liao, Adam Jelley, Eloi Alonso, Florian Laurent, Fredrik Norén, James Swingos, Jan Hünermann, Kent Rollins, Lucas Hosseini, Matthieu Le Cauchois, Maxim Peter, Pim de Witte, Tim Brown, Vincent Micheli, Moritz Böhle, Gabriel de Marmiesse, Viktoriia Sharmanska, Lucia Specia, Michael Black, Patrick Pérez
Read on arXiv →
Key claim
Multiplayer world model maintains stability beyond training duration.
In plain English
In multiplayer environments, existing models often treat other agents as part of the environment, limiting their effectiveness. This paper addresses that gap by introducing a world model that conditions on multiple agents' actions, allowing for more accurate scene changes attribution. The model, trained on extensive gameplay data, maintains stability in its rollouts for significantly longer than its training duration, which is a notable improvement. Builders interested in creating AI for complex, interactive scenarios will find the methodologies and results relevant for enhancing their systems.
Introduces a novel multiplayer world model for dynamic environments.
Demonstrates stable rollouts and systematic evaluations, though lacks extensive baselines.
Deep reliability assessment
The methodology supports the claim that the model can generate stable rollouts in multiplayer environments, but the paper may overclaim its general applicability without testing in diverse game settings beyond Rocket League.
Reproducibility
Yes, the paper provides open source code and dataset links, allowing for reproducibility of the experiments.
Key figure
Figure 1 illustrates the world model's ability to simulate a dynamic game scene from multiple players' viewpoints, maintaining temporal consistency and coherence in the players' perspectives.