SPG-RRT*: A Semantic Perception Guided Path Planning Algorithm for Robotic Arms in Complex 3-D Environments

Abstract

Robotic arm path planning in complex three-dimensional (3D) environments requires balancing feasibility and efficiency. However, classical Rapidly-exploring Random Tree* (RRT*) and its improvements are often inefficient and lack structural awareness of complex 3D environments, making it difficult to balance planning efficiency and path quality. Simultaneously, existing deep learning-based sampling strategies are predominantly designed for two-dimensional environments, which hinders their direct application to 3D path planning problems. To address these issues, this paper integrates a 3D semantic segmentation network with a sampling-based path planning framework, resulting in the Semantic Perception Guided RRT* (SPG-RRT*) algorithm for robotic arm path planning in complex 3D environments. The proposed method utilizes a 3D semantic segmentation model to predict potentially feasible path regions between the start and goal positions from a global view of the environment, thereby biasing the sampling distribution of RRT* in a semantic-aware manner. This targeted guidance elevates the valid sample proportion and promotes faster convergence, without sacrificing path feasibility. Comparative simulations were conducted in three representative 3D environments (random space, narrow space, and trap space). The proposed algorithm substantially improves planning performance over the classical RRT*, particularly in challenging 3D narrow space environments. Compared with RRT*, SPG-RRT* achieves reductions of 62.2% in planning time, 64.2% in the number of search iterations, and 48.1% in the final path length. Furthermore, a real six-degree-of-freedom robotic arm platform was constructed, and experiments in three types of 3D environments demonstrate that the proposed algorithm can successfully perform robotic arm path planning in complex real-world environments. Note to Practitioners—This paper is motivated by the need for efficient and reliable path planning for robotic arms operating in complex, unstructured industrial and agricultural environments, such as those with narrow passages or trap-like obstacles. Traditional sampling-based methods often suffer from slow convergence and high computational costs in such scenarios. This work presents a semantic perception guided RRT* (SPG-RRT*) algorithm that utilizes a deep learning model (ResVox-Mamba) to predict feasible path regions from global environmental maps. This prediction guides the sampling process, significantly reducing invalid attempts. Practitioners can deploy this algorithm on high-degree-of-freedom robotic arms (validated on a real 6-DOF robotic arm) to achieve faster path generation and higher success rates in 3D obstacle-rich settings. However, the current approach is optimized for static environments. Future extensions will address dynamic obstacle avoidance and multi-robot coordination.