Abstract: A vehicle collision avoidance model was established according to adaptive neuro-fuzzy inference system. In order to improve rapidity of convergence, a hybrid algorithm was proposed. For some linear parameters such as consequent parameters, recursive least square algorithm was used to update it. For other nonlinear parameters such as premise parameters, steepest descent method was used to identify it. To get the teaching data to train the model, an experiment was designed. Global position system(GPS) module was adopted in the experiment to get the headway and velocity difference between lead vehicle and following vehicle. Based on the training data obtained by experiment, the model was trained and a control output was simulated. By comparing the simulation result and the experiment data, it shows that the model can simulate manipulation of driver accurately and even more smoothly.