Latent Field Agents: Leveraging Generative Environment Models for High-Dimensional Visual Control

• Jitendra Gupta

  Compunnel Inc.

  Author

This paper proposes a novel architecture for solving high-dimensional visual control tasks without relying on recurrent neural networks (RNNs) or stacked image frames as input. Instead of directly learning policies from raw observations, the proposed framework first constructs a generative model that captures the underlying dynamic processes of the environment. This generative model is then utilized to train the reinforcement learning agent on a compact and informative latent representation of the environment state. By learning a structured latent space, the architecture effectively encodes temporal dependencies and environmental dynamics while reconstructing the complete sequence of observed states. Consequently, the learned representations provide a more meaningful description of the environment than raw visual inputs alone. Furthermore, the probabilistic nature of the generative model enables the framework to capture transition uncertainty, allowing the agent to make more robust decisions under stochastic conditions and improving learning stability. The proposed approach accelerates policy convergence by reducing the complexity of high-dimensional observations while preserving essential dynamic information. The agent operates without any prior knowledge of the environment and learns directly through interaction in a continuous action space using an on-policy actor-critic reinforcement learning algorithm. Experimental results demonstrate that the proposed architecture achieves efficient representation learning, faster convergence, and improved policy performance across challenging visual control environments.

• PDF