基于深度确定性策略梯度自适应DWA算法的移动机器人路径规划

Local Path Planning of Mobile Robot Based on DDPG Adaptive DWA Algorithm

摘要:

针对传统动态窗口法在复杂环境中避障与绕行策略适配性不足的问题,提出基于深度确定性策略梯度(Deep deterministic policy gradient, DDPG)的自适应动态窗口(Adaptive dynamic window approach,ADWA)算法。ADWA在DWA框架中引入DDPG强化学习机制,从结构层面对DWA进行改进,通过策略网络实现代价函数中关键权重参数的动态自调节。具体地,ADWA利用激光雷达数据以及机器人与目标点之间的相对位置信息构建DDPG状态空间,并将Actor网络的输出结果作为DWA中航向角偏差代价项的可变权重,引导运动方向决策。通过与环境的持续交互,算法基于奖励机制不断优化策略函数,实现对不同场景环境的自适应调整。在不同环境中进行了仿真与实验验证,实验结果表明,相对于传统的DWA算法,ADWA算法平均成功到达率为94. 7% ,且平均绕行率仅为2. 7% ,显著提高了机器人在应对各种场景时的灵活性和鲁棒性,并且能够有效提高机器人在复杂环境中的通行效率。

Abstract:

Aiming to address the poor adaptability of the traditional dynamic window approach (DWA) in obstacle avoidance and detour strategy selection within complex and dynamic environments, an adaptive dynamic window approach ( ADWA) enhanced by the deep deterministic policy gradient ( DDPG) algorithm was proposed. The proposed framework integrated a reinforcement learning mechanism into the DWA architecture, allowing online optimization of the motion evaluation process through a policy network that dynamically adjusted the weighting coefficients of key cost function components. A continuous DDPG state space was constructed by using laser radar perception data and the relative position between the robot and the target, while the output of the actor network served as an adaptive weighting factor for the heading deviation cost term, guiding motion decisions in real time. Through continuous interaction with the environment and reward-driven policy optimization, the proposed method autonomously refined its decision-making strategy to adapt to varying obstacle distributions and environmental conditions without manual parameter tuning. Extensive simulations and real-world experiments conducted across diverse and cluttered environments demonstrated that the ADWA achieved an average goal-reaching success rate of 94. 7% and a detour rate of only 2. 7% . Compared with the traditional DWA, the proposed method exhibited superior flexibility, robustness, and navigation efficiency, effectively enhancing the adaptability and overall motion performance of mobile robots in complex and uncertain scenarios. These findings confirmed that integrating DDPG into the DWA framework can provide a scalable and intelligent approach to real-time motion planning for autonomous robotic navigation.