In this paper, the direct way of designing a stable controller for nonlinear system is studied. A framework of learning controller with Lyapunovbased constraint is proposed, which transforms design and analysis of a controller to straightforward way by solving an optimization item with the Lyapunov constraint. A novel way of the global stability guaranteed controller is realized directly. Firstly, the optimization problem subject to Lyapunovbased constraints is formulated, in which the tracking error is the objective function to be minimized. Secondly, the controller combines with PID and feedforward is given in form of neural networks. Finally, the optimization solution of the controller method is analyzed and solved, in which some deep learning technologies are used to enhance the capability of solution. Test results of two simulations of 2 order linear and nonlinear systems demonstrate that the proposed method has high performance in speed of convergence, tracking error and smoothness and amplitude of control output. Results of comparison simulation with backstepping control to the nonlinear system with disturbance, noise, uncertainty of parameters and the difference of reference output demonstrate that the proposed method has high performance in terms of robustness and generalization. Results of simulated physical experiment of onestage rotary inverted pendulum based on VRep and physical testing of singleaxis controlling for quadrotor prove that the method proposed is capable of high precision control and strong disturbance rejection.