Anticipatory force control is a fundamental means by which humans stave off slipping, spilling, and tilting disasters while manipulating objects. this control must often be adapted due to changes in an object's dynamics (e.g. a lighter than expected mug of coffee) or its relation with involved effectors or digits (e.g. lift a mug with three vs. five digits). The neural processes guiding such anticipatory and adaptive control is understudied but presumably operates along multiple time scales, analogous to what has been identified with adaptation in other motor tasks, such as perturbations during reaching. Learning of anticipatory forces must be ultrafast to minimize tilting a visually symmetric object towards its concealed asymmetric center of mass (CoM), but slower when the CoM is explicitly and systematically switched from side to side. Studying the neural substrates of this latter slower learning process with rapid multiband brain imaging, in-scanner kinematics and Bayesian pattern component modelling, we show that CoM-specific pattern distances increase with repeated CoM switching exposures and improved learning. The cerebellum showed the most prominent effects, fitting with the idea that it forms a stored internal model that is used to build and update anticipatory control. CoMspecific pattern distances were present 24 h later, in line with the presence of consolidation effects. Numerous behavioral and computational studies demonstrate that a wide variety of motor behaviors can be adapted to a constantly fluctuating environment and that this learning occurs on multiple time scales 1. For example, prism adaptation, force field adaptation and visuomotor rotation all can be described by a combination of fast and slow learning components. The same general idea of multiple time scales is likely to be operating when people manipulate objects. In this case, when an object's dynamics are mismatched with expectations, the lift forces of the digits adapt extremely rapidly to minimize roll or other aspects of movement error 2. For example, it takes only three trials to successfully lift a visually symmetrical inverted T-shaped object without tilting it towards its concealed asymmetric center of mass (CoM). This is accomplished by an anticipatory partitioning of digit lift forces at lift onset (i.e. more force by the digit closest to the CoM), which generates an appropriate torque counteracting the offset CoM 3. This improvement over just a few trials can be described as an ultrafast form of adaptation 4,5. When the object CoM is then switched from one side to the other (which alters the relation between the involved digits and the object), the movement errors re-emerge, suggesting that ultrafast adaptation is not immediately generalizable to different CoM offsets and appropriate lift forces must be re-acquired 6,7. Critically, this failure to generalize to a new CoM occurs despite a subject's explicit perceptual awareness of the switch (seeing the object being rotated) and an intact knowledge of the CoM (correctly point...