In research of path planning of fruit picking robot the probabilistic roadmap method is adopted to achieve on-line planning with heuristic search arithmetic. At the sampling and searching stage, the density of samples in neighborhood was used as weight to control expansion of roadmap and avoiding over sampling in free space. After finding out best path, we used hierarchical decomposition methods to detect collision, which based on oriented bounding boxes. When picking the fruit, a robot usually needs to round over shelf and pass through narrow region. We simulated these two typical situations with virtual reality technique. The results show that a path planning cycle is less than 0.15 s and meets the request of real-time.