A hydraulic omnidirectional leveling system was designed based on “three-layer frame” and a composite Q learning-BP neural network-PID (QBP-PID) omnidirectional leveling control algorithm was proposed, and the crawler machine was taken as the research object. Firstly, the structural scheme and working principle of the whole omnidirectional leveling machine were given. On this basis, a dynamic model of the entire crawler machine, including the omnidirectional leveling system, was established. Then aiming at the problem that PID control parameters were difficult to be adjusted, an omnidirectional leveling compound QBP-PID controller was established. The PID control parameters were updated in real time through the BP neural network, and the Q-learning algorithm was introduced to update the neural network connection weights online. Results of simulation showed that the leveling time under QBP-PID control was 2.8s for 20° transverse leveling and 3.2s for 25° longitudinal leveling. Compared with PID and BP-PID control, the leveling time was reduced, and no overshoot occurred. In the end, the complete machine test on transverse slope road and longitudinal slope road was carried out. Compared with the simulation results, the errors of transverse and longitudinal leveling time were 0.6s and 0.4s. And the body inclination angle on the horizontal ground was less than 1.5°, which satisfied the leveling performance demand of hilly mountainous agricultural machinery.