Scholarlink

<p>To address the challenges of insufficient dynamic obstacle avoidance capability and limited grasping planning capabilities in complex environments for mobile robotic arms, this paper introduces a hybrid algorithm, COQNLS-TD3(GRU)-PER. This algorithm integrates the modified Constrained Optimization Quasi-Newton Least Squares (COQNLS) method with the enhanced Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, which is further augmented by incorporating the Gated Recurrent Unit (GRU) module and a Prioritized Experience Replay mechanism (PER). Initially, COQNLS planning is designed for the robotic arm to obtain the pre-grasp posture. Subsequently, the problem is defined as a Markov Decision Process involving the design of state space, action space, and reward-penalty functions. Subsequently, the GRU module is integrated into the neural network to process temporal state space features, thus enhancing dynamic obstacle avoidance capabilities. The planning process unfolds in phases: the dynamic obstacle avoidance phase is driven by training with TD3(GRU), and the target planning phase is steered by the pre-grasp posture, complemented by a cooperative guidance mechanism for transitional control outputs. To enhance training sample efficiency and accelerate algorithm convergence, a Prioritized Experience Replay mechanism is incorporated, facilitating the efficient development of an effective policy model for planning. Finally, to validate the efficacy of the proposed algorithm, a three-dimensional experimental scenario is designed for comparative experiments with other algorithms. Experimental results reveal that compared to the traditional TD3 algorithm, the COQNLS-TD3(GRU)-PER algorithm offers substantial improvements in both training efficiency and the control performance of the policy model.</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

Dynamic obstacle avoidance and grasping planning for mobile robotic arm in complex environment based on improved TD3

  • 期刊论文

  • 发表日期：

    2025年6月19日

查看全文
  • Yong Li,
  • Linbing Ke,
  • Chaoxing Zhang

摘要

To address the challenges of insufficient dynamic obstacle avoidance capability and limited grasping planning capabilities in complex environments for mobile robotic arms, this paper introduces a hybrid algorithm, COQNLS-TD3(GRU)-PER. This algorithm integrates the modified Constrained Optimization Quasi-Newton Least Squares (COQNLS) method with the enhanced Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, which is further augmented by incorporating the Gated Recurrent Unit (GRU) module and a Prioritized Experience Replay mechanism (PER). Initially, COQNLS planning is designed for the robotic arm to obtain the pre-grasp posture. Subsequently, the problem is defined as a Markov Decision Process involving the design of state space, action space, and reward-penalty functions. Subsequently, the GRU module is integrated into the neural network to process temporal state space features, thus enhancing dynamic obstacle avoidance capabilities. The planning process unfolds in phases: the dynamic obstacle avoidance phase is driven by training with TD3(GRU), and the target planning phase is steered by the pre-grasp posture, complemented by a cooperative guidance mechanism for transitional control outputs. To enhance training sample efficiency and accelerate algorithm convergence, a Prioritized Experience Replay mechanism is incorporated, facilitating the efficient development of an effective policy model for planning. Finally, to validate the efficacy of the proposed algorithm, a three-dimensional experimental scenario is designed for comparative experiments with other algorithms. Experimental results reveal that compared to the traditional TD3 algorithm, the COQNLS-TD3(GRU)-PER algorithm offers substantial improvements in both training efficiency and the control performance of the policy model.