Hierarchical Decision Transformer

摘要

Sequence models in reinforcement learning require task knowledge to estimate the task policy. This paper presents the hierarchical decision transformer (HDT). HDT is a hierarchical behavior cloning algorithm that improves the performance of transformer methods in imitation learning, improving their robustness to tasks with longer episodes and/or sparse rewards, without requiring task knowledge or user interaction currently present in the state-of-the-art. The high-level mechanism guides the low-level controller through the task by selecting sub-goals for the latter to reach. This sequence replaces the returns-to-go of previous methods, improving its performance overall, especially in tasks with longer episodes and scarcer rewards. We validate our method in multiple tasks of OpenAI Gym, D4RL, and RoboMimic benchmarks. Our method outperforms the baselines in twenty three out of thirty one settings of varied horizons and reward frequencies without prior task knowledge, showing the advantages of the hierarchical model approach for learning from demonstrations using a sequence model. We also evaluate the method on a reaching task on a physical robot.