This experiment investigated the prediction of load force (LF) in impulsive collisions inferred from anticipatory adjustments of grip force (GF) used to stabilise a hand-held object. Subjects used a precision grip to hold the object between thumb and index finger of their right hand and used the arm either: (1) to move the object to produce a collision by hitting the lower end of a pendulum, causing it to swing to one of three target angles, or (2) to hold the object still while receiving a collision produced by the experimenter releasing the pendulum from one of three angles. Visual feedback of the pendulum's trajectory was available in the production task only. In all conditions, subjects increased GF in advance of the collision. In receiving the collision without advance information, subjects set GF levels to the mid-range of the experienced forces. When subjects possessed knowledge about the maximum angle of pendulum swing - either because they were going to produce it or because they were verbally informed - magnitude of the anticipatory-GF magnitude response was scaled to the predicted LF magnitude. Furthermore, GF was scaled to LF with a higher gain when producing compared to receiving the collision. This suggests that updating forward models through a semantic route is not as powerful as when the updating is achieved through the more direct route of dynamic exploration.