There are many repetitive activities in agriculture such as plowing the fields, sowing seeds and sowing fertilizers. Repetitive activities performed using robots will provide better durability and precision. The use of robots in agriculture (robofarming) to follow the specified path (path tracking) can make agricultural activities more efficient. There are many methods of path tracking that can be used. One of the methods used is to use an artificial neural network (ANN). The development of ANN on devices with limited memory and low computing such as microcontrollers will provide a faster response because it reduces the latency of data transfer from the microcontroller to the central computer/cloud. To be able to run on devices with small memory, it is necessary to optimize memory usage and algorithms on the ANN model used. This research implements an artificial neural network-based path tracking on a wheeled robot prototype by controlling the speed on both wheels. Feed-forward ANN is used as an approximator to predict the direction and speed required for the robot to maneuver according to the target point of the map. The ANN model uses the input error heading and the difference between the position of the robot and the trajectory. Tests were carried out using 3 points, 4 points and 5 target points. Based on the test, obtained an average error of ±72.49mm with a standard deviation of 50.40mm and a maximum error of ±236.63mm.