Engineering vehicles operate under high torque, high load, and complex environmental conditions, facing numerous technical challenges. Particularly during the starting phase, the significant slippage of clutch discs significantly impacts the precision of clutch torque control. Therefore, to achieve adaptive start-up control for AMT engineering vehicles, an adaptive control method combining linear quadratic regulator (LQR) and deep neural network was proposed for the AMT start-up process. At the upper level of the control strategy, a constant engine speed strategy was formulated based on different starting intentions, and the LQR was used to obtain the reference speed corresponding to the reference torque of the clutch under different environments. With considering the complexity of the operating environment, a certain range of perturbations was introduced into the vehicle dynamics model to generate a series of reference rotational speed trajectories as the training data set for the deep neural network, and a robust data model offline was obtained. At the lower level of the control strategy, a clutch friction factor adaptive controller was designed to estimate the clutch friction factor in real time. Finally, the effectiveness of the adaptive start control method for engineering vehicles equipped with AMT was verified by simulation tests. The results show that the proposed method has good starting performance under the condition of unknown friction coefficients and can adapt to different starting intentions and driving environments. Compared with the PID controller which does not depend on the mechanism model, it has higher adaptive ability and robustness.