The invention discloses a mars probe onboard autonomous orbit calculation method, which is a calculation method based on radial basis function neural network curve fitting. The method includes the following steps: building a three-layer radial basis function network model; getting orbit prediction data of a mars probe within a period of time according to ground orbit determination, and training the radial basis function network model by taking the data as a training sample; and finally, uploading the trained radial basis function neural network model as an onboard orbit prediction model to the mars probe. There is neither need for establishment of a complex dynamic model on a satellite nor need for ephemeris calculation. The prediction precision is almost equal to the ground orbit prediction precision. The method is also suitable for engineering implementation. The method not only can satisfy the engineering precision constraints of the recursion result, but also can satisfy the constraints of limited resources of onboard computers.