Abstract: Handpicking of waste is a labor-intensive operation with potential safety risks. To tackle these problems, an intelligent robotic system that can perform dirt recognition-sorting integrated operations is specifically designed. The system consists of two parts in design: machine vision algorithm and the robotic manipulators controlling system. For the former, the CornerNet-SqueezeNet model is used, which has a fast operational speed and a remarkable recognition effect while for the latter, parallel manipulator and 3-coordinate manipulator are used for grabbing the gangue. By incorporating the special merits of the high speed and high load-carrying capacity and flexibility of the two manipulators, the queued coal or waste rock pieces can be accurately sorted. Field application of the system at Tangshan Mining Branch of Kailuan Limited Liability Corporation shows that the system works with a dirt recognition time of less than 0.2 s, a recognition rate of 93.05%, a working cycle of manipulators of 1s, and a load-carrying capacity of 15 kg. The coal pieces in a size of 50~300 mm can be sorted out with a success rate of over 90%.