Structured State Representation and Hierarchical Policy Scheduling for Reinforcement Learning-Based Task Planning
- Authors
-
-
Shuncheng Liu
Columbia University, New York, USA
Author
-
- Abstract
-
This study proposes a reinforcement learning-driven hierarchical task planning algorithm to address the challenges faced by agents in complex and dynamic environments, including hierarchical task decomposition, state uncertainty, and limited execution robustness. The proposed method centers on a layered policy architecture, employing a perception encoder to extract environmental features and integrating a task decomposition module with a temporal scheduling mechanism to achieve coordinated optimization between high-level goal planning and low-level action control. Within the overall framework, the perception module transforms environmental information into structured representations, while the temporal feature extractor and graph topology encoder jointly model multi-scale dependencies. The scheduler then generates hierarchical policies based on reinforcement learning signals, dynamically balancing global task objectives and local execution behaviors. To ensure stable policy convergence and behavioral consistency, the algorithm incorporates advantage estimation and entropy regularization mechanisms, achieving controllability and robustness in the decision space through multi-stage optimization constraints. Experimental results demonstrate that the proposed method exhibits superior hierarchical planning capability and execution efficiency across diverse task scenarios, maintaining high task success rates and path consistency under challenging conditions such as high noise, dynamic obstacles, and partial observability. The findings further validate the effectiveness of the reinforcement learning-based hierarchical task planning framework for long-horizon and multi-objective decision-making, providing a general algorithmic foundation and structured modeling approach for applications in autonomous navigation, industrial control, and human-robot collaboration.
- Downloads
- Published
- 2025-11-01
- Issue
- Vol. 5 No. 11 (2025)
- Section
- Articles