Abstract: To address the problems of blind search, obstacle-adjacent paths, and redundant turns in the path planning of feeding robots in narrow indoor marine aquaculture channels, an improved rapidly-exploring random tree (RRT) path planning algorithm with safety cost constraints is proposed. First, a safety cost model is constructed based on the occupancy grid map, and obstacle-distance constraints are incorporated into the node expansion and parent node selection processes, guiding the random tree to expand toward low-risk regions and thereby improving the safety margin of the planned path. Second, goal-biased sampling is employed to enhance the convergence of the random tree toward the target region and reduce ineffective search. Finally, reachability-detection-based path backtracking optimization is performed on the initial path to remove redundant intermediate nodes and reduce the cumulative heading deflection angle. Experiments are conducted on an RK3568 embedded platform to evaluate the algorithm in terms of real-time planning performance, path safety, and path smoothness. The results show that the average planning time of the improved RRT algorithm is 75.5% and 80.8% lower than that of the standard RRT and A* algorithms, respectively. The improved RRT algorithm achieves a safety intrusion rate of 0%. Compared with the standard RRT algorithm, the improved RRT algorithm reduces the average number of path nodes from 38.2 to 8.1 and the cumulative heading deflection angle from 13.9 rad to 3.2 rad.