文献类型：期刊文献

中文题名：未知环境下无人车自主导航探索与地图构建

英文题名：Autonomous Navigation Exploration and Map Construction for Unmanned Vehicles in Unknown Environments

作者：满恂钰[1];刘元盛[2];齐含[3];严超[1];杨茹锦[1]

第一作者：满恂钰

机构：[1]北京联合大学,北京市信息服务工程重点实验室,北京100101;[2]北京联合大学,机器人学院,北京100101;[3]北京联合大学,智慧城市学院,北京100101

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2023

期号：11

起止页码：34-40

中文期刊名：汽车技术

外文期刊名：Automobile Technology

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2023_2024】；

基金：国家重点研发计划项目(2021YFC3001300);国家自然科学基金重点项目合作项目(61931012);北京联合大学高水平孵化项目和新进博士孵化项目(ZK10202208);北京市属高等学校高水平科研创新团队建设支持计划项目(BPHR20220121)。

语种：中文

中文关键词：自主导航探索;长短期记忆网络;深度强化学习;地图构建

外文关键词：Autonomous navigation exploration;Long-Short-Term Memory(LSTM)network;Deep reinforcement learning;Map building

摘要：针对自主导航探索算法易陷入局部区域的问题,提出了融合采样与深度强化学习的探索算法。首先,局部采用长短期记忆(LSTM)网络获得无人车历史位姿信息进而避免重复走向已探索区域;其次,利用深度强化学习输出策略最优的动作并设计奖励函数以激励无人车充分探索未知区域;最后,考虑无人车水平移动因素,通过解非对称旅行商问题(ATSP)生成一条符合其当前姿态的全局探索路径。2000s矿道仿真环境中,所提出的算法相较于无人机自主探索(TARE)算法,探索面积增加346.3m2,总行驶距离减少209.4m;在真实场景试验中,该探索算法用时1014s完成面积为3444.3m2的地下车库探索返回起点,并完成环境地图构建。
For the problem that the autonomous navigation exploration algorithm is easy to fall into the local area,this paper proposed an exploration algorithm combining sampling and deep reinforcement learning.First,the Long-Short-Term Memory(LSTM)network was used locally to obtain the historical pose information of the unmanned vehicle to avoid repeated exploration of the explored area;secondly,the optimal action of the deep reinforcement learning strategy was used to output using deep reinforcement learning and the reward function was designed to encourage the unmanned vehicle to fully explore the unknown area;Finally,the horizontal movement factor of the unmanned vehicle was considered to generate a global exploration path conforming to its current attitude by solving the Asymmetric Travel Salesman Problem(ATSP).In the 2000 s mine tunnel simulation environment,compared with the Technologies for Autonomous Robot Exploration(TARE)algorithm,the proposed algorithm increased the exploration area by 346.3 m2 and reduced the total driving distance by 209.4 m;in the real scene test,the exploration algorithm completed the exploration of the underground garage with an area of 3444.3 m2 and returned to the starting point in 1014 s and built the environment map.