Sensory noise and feedback delay are potential sources of instability and variability for the on-line control of movement. It is commonly assumed that predictions based on internal models allow the CNS to anticipate the consequences of motor actions and protect the movements from uncertainty and instability. However, during motor learning and exposure to unknown dynamics, these predictions can be inaccurate. Therefore a distinct strategy is necessary to preserve movement stability. This study tests the hypothesis that in such situations, subjects adapt the speed and accuracy constraints on the movement, yielding a control policy that is less prone to undesirable variability in the outcome. This hypothesis was tested by asking subjects to hold a manipulandum in precision grip and to perform single-joint, discrete arm rotations during short-term exposure to weightlessness (0 g), where the internal models of the limb dynamics must be updated. Measurements of grip force adjustments indicated that the internal predictions were altered during early exposure to the 0 g condition. Indeed, the grip force/load force coupling reflected that the grip force was less finely tuned to the load-force variations at the beginning of the exposure to the novel gravitational condition. During this learning period, movements were slower with asymmetric velocity profiles and target undershooting. This effect was compared with theoretical results obtained in the context of optimal feedback control, where changing the movement objective can be directly tested by adjusting the cost parameters. The effect on the simulated movements quantitatively supported the hypothesis of a change in cost function during early exposure to a novel environment. The modified optimization criterion reduces the trial-to-trial variability in spite of the fact that noise affects the internal prediction. These observations support the idea that the CNS adjusts the movement objective to stabilize the movement when internal models are uncertain.