This article investigates the improvement of trajectory tracking on an omnidirectional rehabilitative training walker (ORTW) with random shifts in the centre of gravity, control constraints, and touchdown characteristics of omniwheels. Analysis of the touchdown characteristic improved the accuracy of the dynamic ORTW model. The control constraints guarantee the safety of the omniwheels and the ORTW training process. After describing the random centre-of-gravity shifts using Wiener process, an improved stochastic dynamic model was constructed. Under trajectory tracking control with adaptive technology, the ORTW followed the trajectory designed by the rehabilitation therapist. The exponentially practical stability in mean absolute of the tracking-error system was derived from stochastic theory and Markov's inequality. The effectiveness of the proposed methods was confirmed in simulations.