Hierarchical Decision Transformer

TL;DR

This paper introduces a hierarchical decision transformer that improves reinforcement learning from demonstrations by using a high-level sub-goal mechanism, outperforming baselines in diverse tasks without prior task knowledge.

Contribution

The paper proposes a hierarchical sequence model that replaces return-to-go with sub-goal selection, enhancing performance in long-horizon, sparse reward tasks.

Findings

Outperforms baselines in 8 out of 10 tasks

Effective in tasks with long episodes and sparse rewards

No prior task knowledge needed

Abstract

Sequence models in reinforcement learning require task knowledge to estimate the task policy. This paper presents a hierarchical algorithm for learning a sequence model from demonstrations. 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 OpenAIGym, D4RL and RoboMimic benchmarks. Our method outperforms the baselines in eight out of ten tasks 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.