Abstract:

The global aging issue is becoming severe increasingly， and the field of elderly care services is facing a challenge of manpower shortage， urgently requiring the introduction of robot technology with intelligent decision-making capabilities. To solve the autonomous path planning problem of assistive robots under a multi-task mechanism in nursing home scenarios， an improved Soft Actor-Critic （SAC） reinforcement learning decision-making algorithm was proposed. Firstly， an obstacle contour reconstruction method based on virtual circles was introduced， which reduced the complexity of environmental modeling and enhanced radar detection efficiency. Then， to tackle the difficulty of reinforcement learning algorithms in optimizing strategies from scratch when solving complex tasks in a continuous state space， Whale Optimization Algorithm （WOA） was integrated with SAC algorithm to obtain WOA-SAC algorithm. At the same time， by constructing an auxiliary supervision mechanism to provide directional guidance for the learning process， the decision-making capability was improved while the convergence was accelerated significantly. Finally， task planning was conducted on the basis of daily needs of the elderly， and model training was completed in environments composed of fixed tasks with static and dynamic obstacles as well as emergent random tasks. Simulation results demonstrate that compared to the traditional SAC algorithm， WOA-SAC algorithm reduces the average path length by 10.42%， increases the success rate by 6.66%， and decreases the average step size by 29.63%. It can be seen the significant enhancement of WOA-SAC algorithm in the learning efficiency and decision-making capability of SAC algorithm， addressing the autonomous path planning problems in multi-task mechanisms effectively.

Key words: elderly care robot, multi-task path planning, guided learning, Whale Optimization Algorithm (WOA), reinforcement learning

摘要：

全球老龄化问题日益严峻，养老服务领域面临严重人力短缺挑战，亟需引入具有智能决策能力的机器人技术。针对养老院场景下辅助机器人在多任务机制中的自主路径规划问题，提出一种改进的非确定性策略SAC（Soft Actor-Critic）强化学习决策算法。首先，提出基于虚拟圆的障碍物轮廓重构法，在降低环境建模难度的同时提升雷达探测效率；其次，针对强化学习算法在求解连续状态空间内复杂任务时从零进行策略寻优的困难，将鲸鱼优化算法（WOA）与SAC算法结合得到WOA-SAC算法，通过构建辅助监督机制为学习过程提供方向引导，提升决策能力的同时显著提升收敛速度；最后，基于老人的日常需求规划任务，在包含静态障碍、动态障碍的固定任务和突发性随机任务环境中完成模型训练。仿真实验结果表明，与传统SAC算法相比，WOA-SAC算法的平均路径长度缩短了10.42%，成功率提升了6.66%，平均步长减小了29.63%。可见，WOA-SAC算法能够显著提升SAC算法的学习效率和决策能力，并解决多任务机制中的自主路径规划问题。

关键词: 养老机器人, 多任务路径规划, 引导学习, 鲸鱼优化算法, 强化学习