Abstract
We introduce Hierarchical Transformers for Meta-Reinforcement Learning(HTrMRL), a powerful online meta-reinforcement learning approach. HTrMRL aimsto address the challenge of enabling reinforcement learning agents to performeffectively in previously unseen tasks. We demonstrate how past episodes serveas a rich source of information, which our model effectively distills andapplies to new contexts. Our learned algorithm is capable of outperforming theprevious state-of-the-art and provides more efficient meta-training whilesignificantly improving generalization capabilities. Experimental results,obtained across various simulated tasks of the Meta-World Benchmark, indicate asignificant improvement in learning efficiency and adaptability compared to thestate-of-the-art on a variety of tasks. Our approach not only enhances theagent's ability to generalize from limited data but also paves the way for morerobust and versatile AI systems.