Abstract: To address the inefficiency of path computation for multi-degree-of-freedom manipulators in narrow environments, an improved BP-RRT-FN (back propagation-rapidly-exploring random tree*-fast node) obstacle-avoidance trajectory planning algorithm is proposed. To achieve more rapid and accurate convergence in 3D trajectory planning, the node sampling probability is calculated using a distance weighting function. Then, by combining the spherical envelope obstacle model and the axial envelope method, the algorithm calculates an obstacle-free path. Through staged local search, the BP network is trained to predict the number of nodes to be sampled in each local search stage, enabling the algorithm to automatically proceed to the next search stage. To reduce the generation of redundant sampled nodes and improve path optimization efficiency, the FN algorithm is applied to randomly deleting nodes. Simulation results show that the average number of sampled nodes is reduced by 24.96% and 25.30% in two respective scenarios, and the path calculation time is reduced by 6.47 s and 3.87 s, correspondingly. Experimental results on a six-degree-of-freedom manipulator indicate that, during obstacle avoidance movements, the average grasping time and the average search time are reduced by 3.58 s and 3.21 s, respectively. This algorithm can improve the path calculation efficiency of multi-DOF manipulators in multi-obstacle spaces.