The invention discloses an online self-learning method of engine accelerator required torque. According to the online self-learning method of the engine accelerator required torque, a current engine state is identified through an engine state identification module according to a CAN message signal provided by an engine ECU so as to determine starting or stopping of online self-learning of an accelerator required torque MAP learning module in an AMT; input data quality checking and moving window average filtering are conducted on the CAN message signal provided by the engine ECU through an input signal preprocessing module so as to obtain a stable input signal for on-line learning; self-learning state control is conducted through a learning state control module according to the current engine state, the input data quality, coordinates of corresponding data points in a required torque MAP corresponding to the input data, so that switching between different learning states is achieved, and different processing methods correspond to the different learning states. According to the online self-learning method of the engine accelerator required torque, accelerator required torque MAP online self-learning can be conducted during running of a vehicle through the signal provided by the engine, and reliability of AMT gear selecting performance is ensured accordingly.