<p>This work addresses a dynamic crowd-shipping delivery problem with stochastic capacity and available time of crowd shippers. A combination approach of hierarchical reinforcement learning and game theory model is proposed to handle batch-matching, order allocation, and route planning. The upper agent in a reinforce learning model receives a new batch of orders and ODs and then decides to release them immediately or keep for the next cycles to have better chances for matching and routing. It needs to balance between the opportunities to improve the matching and routing and risks of delayed batch-releasing causing lateness. Customer orders are classified to allocate to an appropriate crowd-shipping vehicle by a game theory model. The lower agent in reinforce learning model takes care of routing. The experimental results show that the proposed system outperforms roughly 10% over the benchmarking approach.</p>