Each line segment is supposed to be tangential to the circle at that radius. The centripetal force is directed perpendicular to the tangential velocity. The force can alter the direction of the object.
This seems to be very similar to the instability problems of gradience descent in machine learning. I wonder of some of the mitigations for GD can be applied here.