| 引用本文: | 蒋坤,操菁瑜,柳文章,孙长银,董璐.移动机器人导航与对抗控制的强化学习方法研究[J].控制理论与应用,2025,42(9):1757~1765.[点击复制] |
| JIANG Kun,CAO Jing-yu,LIU Wen-zhang,SUN Chang-yin,DONG Lu.Research on reinforcement learning methods for navigation and adversarial control in mobile robots[J].Control Theory & Applications,2025,42(9):1757~1765.[点击复制] |
|
| 移动机器人导航与对抗控制的强化学习方法研究 |
| Research on reinforcement learning methods for navigation and adversarial control in mobile robots |
| 摘要点击 2814 全文点击 223 投稿时间：2023-03-24 修订日期：2025-02-28 |
| 查看全文 查看/发表评论 下载PDF阅读器 HTML |
| DOI编号 10.7641/CTA.2024.30164 |
| 2025,42(9):1757-1765 |
| 中文关键词 强化学习 移动机器人 导航避障 对抗策略 |
| 英文关键词 reinforcement learning mobile robot navigation and obstacle avoidance confrontation policy |
| 基金项目 国家自然科学基金项目(62236002,61921004,62173251)资助. |
|
| 中文摘要 |
| 传统机器人导航和决策方法过度依赖于高精度地图的构建,且难以适应动态复杂的应用场景.此外,现有
基于机器学习算法的导航和控制方法在真实系统中存在着泛化和迁移能力不理想的缺陷.针对上述问题,本文提出
了一种基于多模信息融合和强化学习框架的移动机器人导航和实时对抗方法.首先,利用不同类型的信息预处理
模块对机器人采集到的RGB图像、激光雷达数据和其他向量信息进行预处理并融合,实现了机器人对环境的全面
感知. 然后,基于动作网络直接输出机器人运动控制指令,完成了无模型情况下对机器人端到端的控制.进一步的,
在仿真系统中充分考虑现实环境中的噪声和动态因素,利用迁移到实体机器人上的测试数据对模型进行微调和修
正. 最后,在仿真环境和真实搭建的环境中开展不同难度导航和实时对抗任务的实验,验证了所提出的基于强化学
习的机器人导航和实时对抗策略的有效性. |
| 英文摘要 |
| The traditional robot navigation and decision-making methods rely heavily on the construction of high
precision maps, and are difficult to adapt to dynamic and complex application scenarios. In addition, the existing navigation
and control methods based on machine learning algorithms have the defects of unsatisfactory generalization and transfer
ability in real systems. To solve the above problems, a mobile robot navigation and real-time confrontation method based
on multimodal information fusion and reinforcement learning framework is proposed in this paper. First of all, various
information preprocessing modules are used to preprocess and fuse the RGB images, LiDAR data and other vector infor
mation collected by the robot, so as to realize the robot’s comprehensive perception of the environment. Then, the system
directly outputs the motion control commands of the robot based on the action network, allowing for the end-to-end control
of the mobile robot without a model. Furthermore, the noise and dynamic factors in the real environment are fully consid
ered in the simulation system, and the model is fine-tuned and corrected by using the test data migrated to the real robot.
Finally, experiments on navigation and real-time confrontation tasks of different difficulties are carried out in the simulation
environment and the real environment, and the effectiveness of the proposed robot navigation and real-time confrontation
method based on reinforcement learning is verified. |
|
|
|
|
|