Multi-Perspective Self-Supervised Learning-Guided Goal-Conditioned Reinforcement Learning

Multi-Perspective Self-Supervised Learning-Guided Goal-Conditioned Reinforcement Learning

연구 분야: Artificial Intelligence

학회: CIBDA '25: Proceedings of the 2025 6th International Conference on Computer Information and Big Data Applications

Goal-conditioned reinforcement learning (GCRL) enables agents to accomplish specific tasks and demonstrates broad application prospects across multiple domains. However, in complex environments, task diversity significantly impacts policy training efficiency. To address this challenge, this paper proposes a multi-perspective self-supervised learning-guided goal-conditioned RL method. The approach introduces a goal relabeling mechanism to construct expert demonstrations from both exploration and exploitation perspectives. By leveraging behavioral cloning to learn distinct expert policies, the method enhances the learning efficiency of goal-conditioned RL. Experimental results on robotic manipulation tasks validate that our method effectively improves task success rates and convergence speed, demonstrating its feasibility in enhancing goal-conditioned RL performance in complex environments.