New Framework Boosts Meta-RL Adaptation and Sample Efficiency

This research introduces a novel framework designed to enhance meta-reinforcement learning (Meta-RL) by addressing a common limitation: the tight coupling of task inference with agent-specific control mechanisms. The proposed method aims to separate these concerns, enabling more efficient knowledge transfer and adaptation across different robotic embodiments. The framework operates by first learning abstract, task-level knowledge using a simplified agent model. This knowledge, organized by a Bayesian non-parametric prior, is then transferred to more complex, heterogeneous agents. A key innovation is the semantic-magnitude interface and a lightweight temporal adaptor, which translate the learned meta-knowledge into actionable subgoals for various low-level controllers. Experimental results demonstrate substantial improvements in performance, with the framework reducing final-step tracking error by a significant margin compared to existing state-of-the-art baselines. Furthermore, it achieves comparable deployment performance while requiring considerably less interaction data, highlighting its efficiency and potential for broader application in robotics and AI.