Select-to-Act: Hierarchical Reinforcement Learning via Adaptive Language Guidance

The paper presents a novel framework called Hierarchical Reinforcement Learning with Language Instructions (HRLLI), which enhances sample efficiency in RL by using dynamically selectable natural-language instructions as guidance. HRLLI employs a two-level policy structure within a Select-to-Act paradigm, where a high-level policy selects relevant instruction pieces based on the current state, while a low-level policy executes actions conditioned on this guidance. Experimental results on the RTFM benchmark indicate that HRLLI significantly outperforms existing instruction-conditioned RL methods, highlighting the importance of adaptive instruction selection in complex decision-making environments.