To address the low efficiency, poor adaptability, and slow convergence of traditional path planning algorithms, this paper proposes a robotic arm path planning method integrating collision recovery with dynamic Gaussian sampling. The method employs a mixed Gaussian sampling strategy that dynamically adjusts weights for four components: Collision recovery, obstacle avoidance, goal-biasing, and uniform exploration. By evaluating real-time states such as consecutive collisions and goal proximity, it strikes a dynamic balance between global exploration and local recovery, significantly enhancing sampling efficiency in complex environments. The collision recovery component further enhands performance by narrowing the sampling range to the feasible space based on collision frequency and position. Furthermore, a multi-level adaptive step-size extension mechanism is introduced, which adjusts to local environments by considering collision feedback, obstacle density, and global planning progress. This reduces repeated collisions in narrow passages and inefficient exploration in open areas. Finally, a joint optimization strategy combining a K-dimensional tree (KD-Tree), redundancy removal, and an early-stopping mechanism boosts computational speed and convergence. Simulation results indicate that compared with dynamic variable sampling area RRT (DVSA-RRT), the proposed method reduces planning time by 87.41%, sampling nodes by 91.01%, and path length by 7.46% in narrow passages, while generating safe, smooth, and asymptotically optimal paths. Moreover, physical experiments verify its effectiveness in actual scenarios, ensuring the planned paths meet surface roughness measurement requirements.