English

From Pixels to Cooperation Multi Agent Reinforcement Learning based on Multimodal World Models

Multiagent Systems 2025-11-12 v2

Abstract

Learning cooperative multi-agent policies directly from high-dimensional, multimodal sensory inputs like pixels and audio (from pixels) is notoriously sample-inefficient. Model-free Multi-Agent Reinforcement Learning (MARL) algorithms struggle with the joint challenge of representation learning, partial observability, and credit assignment. To address this, we propose a novel framework based on a shared, generative Multimodal World Model (MWM). Our MWM is trained to learn a compressed latent representation of the environment's dynamics by fusing distributed, multimodal observations from all agents using a scalable attention-based mechanism. Subsequently, we leverage this learned MWM as a fast, "imagined" simulator to train cooperative MARL policies (e.g., MAPPO) entirely within its latent space, decoupling representation learning from policy learning. We introduce a new set of challenging multimodal, multi-agent benchmarks built on a 3D physics simulator. Our experiments demonstrate that our MWM-MARL framework achieves orders-of-magnitude greater sample efficiency compared to state-of-the-art model-free MARL baselines. We further show that our proposed multimodal fusion is essential for task success in environments with sensory asymmetry and that our architecture provides superior robustness to sensor-dropout, a critical feature for real-world deployment.

Keywords

Cite

@article{arxiv.2511.01310,
  title  = {From Pixels to Cooperation Multi Agent Reinforcement Learning based on Multimodal World Models},
  author = {Sureyya Akin and Kavita Srivastava and Prateek B. Kapoor and Pradeep G. Sethi and Sunita Q. Patel and Rahu Srivastava},
  journal= {arXiv preprint arXiv:2511.01310},
  year   = {2025}
}

Comments

We have identified critical issues in the code implementation that severely deviate from Algorithm 1, invalidating all experimental results and conclusions. Despite exhaustive efforts to correct these issues, we find they fundamentally undermine the paper's core claims. To uphold academic integrity and prevent misinformation, we are withdrawing this manuscript

R2 v1 2026-07-01T07:18:47.807Z