Scholarlink

Hierarchical reinforcement learning addresses long-horizon tasks by decomposing control into high-level planning and low-level skill execution. However, training both levels jointly remains challenging due to exploration inefficiency and limited subpolicy reuse across environments. We propose an HRL framework that integrates curriculum learning and subpolicy transfer for navigation tasks. A high-level policy selects symbolic subgoals, while a low-level policy executes goal-conditioned behaviors via a shared input abstraction. This abstraction decouples the low-level policy from environment-specific encodings, enabling skill reuse across tasks. We apply curriculum learning to both levels: the low-level policy is trained on environments of increasing difficulty, and the high-level policy learns progressively longer tasks. We evaluate our approach in custom GridNav environments and the symbolic Taxi-v3 domain. Results show that the curriculum-trained low-level policy transfers effectively, and high-level planning benefits from staged learning and reusable primitives. The framework supports sample-efficient, modular learning suitable for scalable multi-task navigation.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hierarchical Reinforcement Learning with Curriculum Learning and Subpolicy Transfer in Navigation Environments

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Lingfei Wang,
  • Yiting Huang

摘要

Hierarchical reinforcement learning addresses long-horizon tasks by decomposing control into high-level planning and low-level skill execution. However, training both levels jointly remains challenging due to exploration inefficiency and limited subpolicy reuse across environments. We propose an HRL framework that integrates curriculum learning and subpolicy transfer for navigation tasks. A high-level policy selects symbolic subgoals, while a low-level policy executes goal-conditioned behaviors via a shared input abstraction. This abstraction decouples the low-level policy from environment-specific encodings, enabling skill reuse across tasks. We apply curriculum learning to both levels: the low-level policy is trained on environments of increasing difficulty, and the high-level policy learns progressively longer tasks. We evaluate our approach in custom GridNav environments and the symbolic Taxi-v3 domain. Results show that the curriculum-trained low-level policy transfers effectively, and high-level planning benefits from staged learning and reusable primitives. The framework supports sample-efficient, modular learning suitable for scalable multi-task navigation.