Urban metros have become the foremost public transit to modern cities, carrying millions of daily rides. As travel efficiency matters to the work productivity of the city, shortening passengers' travel time for metros is therefore a pressing need, which can bring substantial economic benefits. In this paper, we study a fine-grained, safe, and energy-efficient strategy to improve the efficiency of metro systems by dynamically scheduling dwell time for trains. However, developing such a strategy is very challenging because of three aspects: 1) The objective of optimizing the average travel time of passengers is complex, as it needs to properly balance passengers' waiting time at platforms and journey time on trains, as well as considering long-term impacts on the whole metro system; 2) Capturing dynamic spatio-temporal (ST) correlations of incoming passengers for metro stations is difficult; and 3) For each train, the dwell time scheduling is affected by other trains on the same metro line, which is not easy to measure. To tackle these challenges, we propose a novel deep neural network, entitled AutoDwell. Specifically, AutoDwell optimizes the long-term rewards of dwell time settings in terms of passengers' waiting time at platforms and journey time on trains by a reinforcement learning framework. Next, AutoDwell employs gated recurrent units and graph attention networks to extract the ST correlations of the passenger flows among metro stations. In addition, attention mechanisms are leveraged in AutoDwell for capturing the interactions between the trains on the same metro line. Extensive experiments on two real-world datasets collected from Beijing and Hangzhou, China, demonstrate the superior performance of AutoDwell over several baselines, capable of saving passengers' overall travel time. In particular, the model can shorten the waiting time by at least 9%, which can boost passengers' experience significantly.