Abstract: Autonomous agricultural vehicle (AAV) is believed to play a significant role in smart agriculture. In future periods, AAV will still be dominated by fuel engines, and the transmission technology will be upgraded from manual shifting to power shifting or CVT (Continuously Variable Transmission) simultaneously for which to improve the power and economy. In response to the demand of AAV, this study devotes to developing an autonomous driving and operation system based on ROS (Robot Operating System) and CVT tractor. The proposed system includes safety, planning, control and CAN bus communication modules. To do that, we integrated and deployed hardware on CVT tractor, designed a CAN bus protocol and implemented data structure for communication between ROS and CAN bus. A close-range anti-collision capability of the tractor is realized based on radar, and a lateral controller based on PID algorithm is adopted. To further validation, a corn sowing experiment was carried out in Miyun District, Beijing, with a total operating area of 2.4 hectares. This research takes DF2204 CVT tractor as the test platform according to the working requirements of autonomous tractors. We also designed the hardware platform of AAV, which was divided into computing layer, sensing layer and actuation layer. Based on the idea of modularization and hierarchy, a software architecture that meets the job requirements was developed with ROS as the middleware, which includes localization, planning, control, CAN communication, and safety modules. And a CAN bus protocol to meet the needs of vehicle control was developed. According to the control characteristics and operation requirements of the CVT tractor, a lateral control module and a longitudinal velocity control module were designed. We counted 22 straight working paths with a total working time of 5 943 s (excluding supplemental seeds and fertilizers), of which the total working time of the straight working stage was 4 037 s and the time of the U-turn stage was 1 906 s, approximately 32% of the total working time. The efficiency of our system was 1.33 hm2/h, and the experimental result shows that the communication node could meet the communication requirements of 50 Hz.The target speed was set to 3 km/h when the tractor turns around, while the harrow was stopped and the seeder was lifted. The engine load was lower when the tractor makes a U-turn, and the torque percentage fluctuates between 10%-30%. After several seconds the engine load continued to increase, and the torque output percentage remained above 80%, with a maximum value of 94%. The engine and transmission will perform a torque reserve in order to meet the instantaneous torque demand (such as increased resistance, climbing, etc.). When the engine revs is higher than the revs corresponding to the torque peak, the increased load will cause the engine to overload, thereby reducing the revs and ego velocity, so as to output more torque. The average lateral error is 2.96 cm and the navigation error was 11.69 cm. The velocity RMSE is 0.98 km/h and MAE was 0.68 km/h. The steering angle RMSE was 1.91°, and MAE was 1.47°. This research shows that the tractor based on wire control technology and CVT tractor can fit the needs of autonomous agricultural vehicles, and the control and planning modules can meet sowing operations. This research could provide a reference for the unmanned upgrading of CVT tractors, and improve the intelligent level and operation efficiency of agricultural machinery.