The invention relates to a robot multi-joint self-adaptation compensation method based on a perceptron model and a stabilizing device. The method comprises the steps that firstly, based on a robot inertial sensor and a gait generator, error data are obtained; secondly, the error data are input into the pre-built perceptron model, the perceptron model updates a network weighting parameter based onan associative learning strategy, and compensation value is calculated and output; and thirdly, based on the compensation value, a robot is subjected to motion compensation, and the compensation valueoutput by the perceptron model comprises hip joint compensation value, knee joint compensation value and ankle joint compensation joint. Compared with the prior art, the associative learning strategyis adopted, the network weighting parameter is updated through a supervisory Hebb learning rule, robustness is improved, self-adaptation control is achieved, compensation amount is dispersed to all joints, such as the ankle joint, the knee joint and the hip joint, of a leg, the joint loads are reduced, and the service life of a robot is prolonged.