An adaptive control method and system for an upper limb rehabilitation robot based on a game theory and surface Electromyography (sEMG) is disclosed. A movement trajectory that a robot is controlled to run within a training time is designed during subject operation. An sEMG-based Back Propagation Neural Network (BPNN) muscle force estimation model establishes a nonlinear dynamic relationship between an sEMG signal and end force by constructing a three-layer neural network. A human-computer interaction system is analyzed by the game theory principle, and a role of the robot is deduced. The control rate between the robot and a subject is updated by Nash equilibrium, and adaptive weight factors of the robot and the subject are determined. The robot adaptively adjusts the training mode thereof according to a movement intention of the subject during operation and a weight coefficient obtained by the game theory principle.