Patients with chronic bilateral vestibular loss have large gaze variability and experience disturbing oscillopsia, which impacts physical and social functioning, and quality of life. Gaze variability and oscillopsia in these patients are attributed to a deficient vestibulo-ocular reflex, i.e. impaired online feedback motor control. Here, we assessed whether the lack of vestibular input also affects feed-forward motor learning, i.e. the ability to choose optimal movement parameters that minimize variability during active movements such as combined eye-head gaze shifts. A failure to learn from practice and reshape feed-forward motor commands in response to sensory error signals to achieve appropriate movements has been proposed to explain dysmetric gaze shifts in patients with cerebellar ataxia. We, therefore, assessed the differential roles of both sensory vestibular information and the cerebellum in choosing optimal movement kinematics. We have previously shown that, in the course of several gaze shifts, healthy subjects adjust the motor command to minimize endpoint variability also when movements are experimentally altered by an increase in the head moment of inertia. Here, we increased the head inertia in five patients with chronic complete bilateral vestibular loss (aged 45.4 ± 7.1 years, mean ± standard deviation), nine patients with cerebellar ataxia (aged 56.7 ± 12.6 years), and 10 healthy control subjects (aged 39.7 ± 6.3 years) while they performed large (75° and 80°) horizontal gaze shifts towards briefly flashed targets in darkness and, using our previous optimal control model, compared their gaze shift parameters to the expected optimal movements with increased head inertia. Patients with chronic bilateral vestibular loss failed to update any of the gaze shift parameters to the new optimum with increased head inertia. Consequently, they displayed highly variable, suboptimal gaze shifts. Patients with cerebellar ataxia updated some movement parameters to serve the minimum variance optimality principle but inaccurately undershot the target leading to an average gaze error of 11.4 ± 2.0°. Thus, vestibulopathy leads to gaze variability not only as a result of deficient online gaze control but also a failure in motor learning because of missing error signals. Patients with cerebellar ataxia in our setting can learn from practice—similar to recent findings in reaching movements—and reshape feed-forward motor commands to decrease variability. However, they compromise optimality with inaccurately short movements. The importance of vestibular information for motor learning implies that patients with incomplete bilateral vestibulopathy, and patients with cerebellar ataxia, should be advised to actively move their head whenever appropriate. This way, sensory error signals can be used to shape the motor command and optimize gaze shifts trial-by-trial.