Scholarlink

<p>Robot path planning is an essential requirement for a robot’s autonomous and collision free movement in the environment. Robots are used to solve human problems through artificial intelligence. In this paper, a Deep Reinforcement Learning (DRL) based end-to-end maples navigation system is designed where an agent starts navigation from an initial position and moves towards the goal position by avoiding obstacles without prior knowledge of the environment. Specifically, this article proposes an obstacle avoidance and navigation policy for a mobile robot with a monocular camera using DRL algorithms. The proposed approach combines two tasks: first, calculating a depth map for obstacle detection and avoidance, and second, determining navigation control commands to reach a target point. For the first step, a fine tuned fully convolutional residual network, a supervised learning approach inspired by Resnet-50 based depth estimation algorithm is used. And, for the second step, deep reinforcement learning based algorithms, such as Deep Q-network (DQN) and its variants Double Deep Q-network (DDQN) and Dueling Double Deep Q-network (D3QN) is employed with a well-designed reward function. Additionally, the problem of sparse reward is handled during designing a hybrid reward allocation function which considers multiple aspects and their associated significance for the navigation. Moreover, we employ the Clustering Based Distribution Factor (CBDF) to lead the robot in all directions to explore the environment to search for the goal point. To validate the performance and robustness of the proposed approach, we carried out the training and testing using robot operating system (ROS) and Gazebo simulator on a pioneer P3DX robot in three different environments. The experiments show that the agent’s behaviour is affected by certain factors such as the complexity of the environments and the number of zones in its region of interest (ROI) area. The results indicate that the D3QN algorithm outperforms the other two algorithms in terms of faster learning and convergence time.</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 reward shaping approach used for end-to-end mapless navigation using deep reinforcement learning algorithm

  • 期刊论文

  • 发表日期：

    2025年9月15日

查看全文
  • Anu Priya,
  • Ritu Tiwari,
  • Pooja Agrawal,
  • Sushant Kumar

摘要

Robot path planning is an essential requirement for a robot’s autonomous and collision free movement in the environment. Robots are used to solve human problems through artificial intelligence. In this paper, a Deep Reinforcement Learning (DRL) based end-to-end maples navigation system is designed where an agent starts navigation from an initial position and moves towards the goal position by avoiding obstacles without prior knowledge of the environment. Specifically, this article proposes an obstacle avoidance and navigation policy for a mobile robot with a monocular camera using DRL algorithms. The proposed approach combines two tasks: first, calculating a depth map for obstacle detection and avoidance, and second, determining navigation control commands to reach a target point. For the first step, a fine tuned fully convolutional residual network, a supervised learning approach inspired by Resnet-50 based depth estimation algorithm is used. And, for the second step, deep reinforcement learning based algorithms, such as Deep Q-network (DQN) and its variants Double Deep Q-network (DDQN) and Dueling Double Deep Q-network (D3QN) is employed with a well-designed reward function. Additionally, the problem of sparse reward is handled during designing a hybrid reward allocation function which considers multiple aspects and their associated significance for the navigation. Moreover, we employ the Clustering Based Distribution Factor (CBDF) to lead the robot in all directions to explore the environment to search for the goal point. To validate the performance and robustness of the proposed approach, we carried out the training and testing using robot operating system (ROS) and Gazebo simulator on a pioneer P3DX robot in three different environments. The experiments show that the agent’s behaviour is affected by certain factors such as the complexity of the environments and the number of zones in its region of interest (ROI) area. The results indicate that the D3QN algorithm outperforms the other two algorithms in terms of faster learning and convergence time.