Scholarlink

<p>For multiple sequential targets, standard path-planning methods need to recompute the complete path for each target allocation independently, thus increasing the computational cost. Even the reward-based learning methods from dynamic programming and reinforcement learning (RL) require an independent re-computation for every new target allocation in the same environment. To solve such multi-goal navigation problems, this paper proposes an RL-based modified algorithm that adopts a region-based approach to reduce the computational cost. The proposed algorithm avoids recomputing the state values and the corresponding policy for every new target in the environment. Instead, it buffers the initially learned state values and corresponding policy based on the various regions of the cluttered environment for later use rather than recomputing them again. The buffered data are improved intermittently based on the maximum Euclidean distance between the start and the goal states. This region-based approach reduces the computational effort for multiple sequential targets from the same or different start locations. Moreover, the proposed approach substantially reduces the path planning time and thus improves the robot's response to interact efficiently in the environment. For the robotic agent, a reward-based two-dimensional cluttered environment is modeled as a Markov decision process (MDP). Numerical simulations are performed to validate theoretical developments for a nonholonomic autonomous ground vehicle (AGV). The results verify the effectiveness of the proposed RL approach as compared to the conventional path-planning algorithms.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A reinforcement learning approach for multi-goal motion planning of autonomous ground vehicles in cluttered environments

  • 期刊论文

  • 发表日期：

    2025年7月5日

查看全文
  • Ghulam Farid,
  • Lanyong Zhang,
  • Talha Younas,
  • Sohaib Tahir,
  • Ahsan Elahi,
  • Muhammad Usman

摘要

For multiple sequential targets, standard path-planning methods need to recompute the complete path for each target allocation independently, thus increasing the computational cost. Even the reward-based learning methods from dynamic programming and reinforcement learning (RL) require an independent re-computation for every new target allocation in the same environment. To solve such multi-goal navigation problems, this paper proposes an RL-based modified algorithm that adopts a region-based approach to reduce the computational cost. The proposed algorithm avoids recomputing the state values and the corresponding policy for every new target in the environment. Instead, it buffers the initially learned state values and corresponding policy based on the various regions of the cluttered environment for later use rather than recomputing them again. The buffered data are improved intermittently based on the maximum Euclidean distance between the start and the goal states. This region-based approach reduces the computational effort for multiple sequential targets from the same or different start locations. Moreover, the proposed approach substantially reduces the path planning time and thus improves the robot's response to interact efficiently in the environment. For the robotic agent, a reward-based two-dimensional cluttered environment is modeled as a Markov decision process (MDP). Numerical simulations are performed to validate theoretical developments for a nonholonomic autonomous ground vehicle (AGV). The results verify the effectiveness of the proposed RL approach as compared to the conventional path-planning algorithms.