LatentMAS

LatentMAS is an advanced framework for multi-agent reinforcement learning (MARL) that uses latent variable modeling to bridge perception and decision-making in environments where agents must coordinate under uncertainty. It provides mechanisms for agents to learn high-level latent representations of states, which simplifies complex sensory inputs into compact, actionable embeddings that facilitate both individual policy learning and inter-agent coordination. Using this latent space, the framework enables Multi-Agent Systems (MAS) to scale more effectively in environments with high dimensionality — such as robotics, simulated physics tasks, and strategic games — by reducing redundant learning burdens and focusing agent exploration. LatentMAS also implements centralized training with decentralized execution, letting agents share learned representations during training while operating autonomously at inference time.