A Goal-Conditioned Reinforcement Learning Algorithm with Environment Modeling

Goal-conditioned Reinforcement Learning (GcRL) has achieved remarkable success in navigating towards goals in recent years. However, learning efficiency and generalization ability remain challenging issues when dealing with uncertain motion patterns of dynamic objects in the environment. To address these issues, existing model-based GcRL algorithms with the environmental prediction have become the main stream. However, these methods are limited for capturing uncertain motion patterns. In fact, humans will consider explicit knowledge, including environmental prediction and policy priors, when making decisions. To improve navigation ability in uncertain environments, we present a novel approach that integrates the policy priors model, which is utilized to infer navigational direction based on the agent's successful historical trajectories, into the model-based GcRL algorithms. Our experimental results demonstrate that our approach is more reliable in navigating towards goals, and outperforms competitive prior works in challenging environments.