Abstract: To address the issues of high randomness, low search efficiency, difficulty in connection, and poor path quality in the traditional bi-directional rapidly-exploring random tree (Bi-RRT) algorithm, an improved Gaussian-biased BiRRT (GB-BiRRT) algorithm is proposed. This algorithm integrates adaptive Gaussian layered sampling (AGLS) and a dynamic target probability bias strategy. A multi-level sampling weight function is constructed based on the distribution of obstacles to constrain the sampling of random trees in invalid areas, concentrating the sampling points in high-probability path regions, thereby improving search efficiency. Additionally, by improving dynamic bias probability adjustment based on the path tree depth, the expansion of the two trees maintains broad search capability while enhancing directional guidance, thus accelerating convergence. Furthermore, a dynamic step size strategy based on environmental complexity is proposed to reduce collision risk. Path optimization is performed through redundant point removal based on collision detection and B-spline curves. Experimental results show that, compared to the traditional Bi-RRT algorithm, the GB-BiRRT reduces path planning time, path length, and the number of sampling points by up to 53.2%, 10.6%, and 46.6%, respectively, in a 2D environment, and by 91.0%, 30.1%, and 74.3%, respectively, in a 3D environment. In verification on a sorting robotic arm platform, the path planning time and path length are reduced by 18.1% and 10.7%, respectively.