Knowledge Reutilization in Meta-Reinforcement Learning

The paper introduces a meta-knowledge reutilization framework for meta-reinforcement learning that decouples task inference from embodiment-specific control, enhancing sample efficiency and facilitating knowledge transfer across heterogeneous agents. It employs a Bayesian non-parametric prior for organizing latent task modes and a high-level policy for task-level guidance, achieving a significant reduction in final-step tracking error (94.75% - 99.79%) compared to state-of-the-art methods, while requiring only 23.8% of the interaction data for comparable performance. This framework is crucial for practitioners as it enables more efficient training and deployment of diverse RL agents by leveraging shared task knowledge.