During object manipulation, both predictive feedforward and reactive feedback mechanisms are available to adjust grip force (GF) levels to compensate for the destabilizing effects of load force changes. During collisions, load force increases impulsively (< 20 ms). Thus, only predictive control of GF can be used to ensure grasp stabilization. A collision paradigm is here used to investigate the effects of practice and vision on the efficiency of the predictive control of GF. Subjects actively produced or received an imposed collision with a pendulum. Subjects were more efficient (used smaller GF for identical loads) when producing than when receiving the collisions. Effects of practice were evident in the active producing task only, with GF levels reducing over repetitions, suggesting that sensorimotor memory for the task was used to adjust GF more efficiently. With imposed collisions, GF levels did not reduce with repetition, which suggests that a direct relation between motor action and sensory feedback may be necessary to improve efficiency. Nevertheless, in this condition GF was lower with visual feedback, indicating potential for more efficient grip possibly associated with subjects degree of confidence. We discuss the implications of these results for accounts of the predictive and the reactive control of movement.