Combining Neural Network and RRT*: A Novel Path Planning Method for Hyper‐Redundant Manipulators With 2 N  + 1 DOF

ABSTRACT Path planning has long been a research focus for hyper‐redundant manipulators. However, in complex environments, the process often entails substantial computational costs. This issue is compounded by kinematic constraints. Therefore, achieving efficient path planning under these conditions remains a significant challenge. To address this issue, this paper proposes a novel path planning method for hyper‐redundant manipulators with 2n + 1 degrees of freedom (DOF), which integrates a Neural Network model with the Rapidly‐exploring Random Tree Star (RRT*) algorithm. First, a general kinematic model of the manipulator is established. Based on this model, the relationship between the manipulator's joint angles and the path corner angle is analyzed. This relationship is then incorporated into the RRT* algorithm to ensure that the generated paths comply with the kinematic constraints of the manipulator. Next, a Neural Network model is developed and trained using a dataset generated by the improved RRT* algorithm. The trained model predicts path points toward a designated goal. To address potential non‐ideal path points in the predictions, a modification strategy combining the Artificial Potential Field (APF) method and a spatial meshing strategy is introduced. This strategy employs a dual adjustment mechanism to reposition non‐ideal path points, enabling obstacle avoidance and mitigating the issue of bilateral distribution in the predicted path points. The predicted path points are then incorporated into the sampling process of the improved RRT* algorithm to guide tree expansion, thereby accelerating the path search. Additionally, the APF method is integrated into the search mechanism to further enhance planning efficiency. A series of experiments was conducted to evaluate the performance of the proposed method. The results demonstrate that the manipulator successfully reaches designated targets along the generated paths while avoiding obstacles and adhering to kinematic constraints. Compared to the RRT and RRT* algorithms, the proposed method shows superior overall performance, particularly in computation time, where it achieves a reduction of over 50% compared to RRT*. Prototype experiments further confirm the feasibility and safety of the paths generated by the proposed approach.