Robotic shepherding problem considers the control and navigation of a groupof coherent agents (e.g., a flock of bird or a fleet of drones) through themotion of an external robot, called shepherd. Machine learning based methodshave successfully solved this problem in an empty environment with noobstacles. Rule-based methods, on the other hand, can handle more complexscenarios in which environments are cluttered with obstacles and allow multipleshepherds to work collaboratively. However, these rule-based methods arefragile due to the difficulty in defining a comprehensive set of rules that canhandle all possible cases. To overcome these limitations, we propose the firstknown learning-based method that can herd agents amongst obstacles. By usingdeep reinforcement learning techniques combined with the probabilisticroadmaps, we train a shepherding model using noisy but controlled environmentaland behavioral parameters. Our experimental results show that the proposedmethod is robust, namely, it is insensitive to the uncertainties originatedfrom both environmental and behavioral models. Consequently, the proposedmethod has a higher success rate, shorter completion time and path length thanthe rule-based behavioral methods have. These advantages are particularlyprominent in more challenging scenarios involving more difficult groups andstrenuous passages.