Abstract
To address unstable value estimation and insufficient use of temporal information by the soft Actor–Critic (SAC) algorithm in dynamic environments, we propose an improved SAC-based path-planning method that integrates distributional value learning and long short-term memory (LSTM). First, the critic models the return distribution via quantile regression to reduce the variance of value estimates and improve the stability of policy evaluation. Second, an LSTM encodes historical obstacle observations to enhance prediction of environmental dynamics and decision-making. Through comparative experiments in dynamic scenarios on the Gazebo simulation platform, the proposed method outperforms LSTM-enhanced DDPG, vanilla SAC, distributional SAC, and LSTM-SAC in average path length, average steps, and average planning time. Moreover, real-world experiments further verify the feasibility and robustness of the approach.
Funding Information
  • Guangxi Key Research and Development Program (AB25069274)
  • Key Laboratory of AI and Information Processing (Hechi University), Education Department of Guangxi Zhuang Autonomous Region (2022GXZDSY003)
  • Guangxi Science and Technology Major Program (AA23023015)

This publication has 36 references indexed in Scilit: