When subjects learn motor tasks under novel visuomotor conditions variations in sensory input during training facilitate adaptive generalization. We tested the hypotheses that training with multiple sensory input variations is more effective than a single or no variation and that training must include critical features of the criterion task. Normal adults were pre- and post-tested on an obstacle avoidance task while wearing visual distortion lenses after treadmill walking (Experiment 1), or balance training (Experiment 2). Subjects were randomized to training groups in which they wore either: (1) three different visual distortion lenses, (2) a single pair of visual distortion lenses, or (3) sham lenses. Post-tests were done while wearing novel lenses. In Experiment 1 subjects who trained with multiple lenses adapted better than single or sham lens groups. The single lens-training group with magnifying lenses adapted better than the other single lens groups. In Experiment 2, training for dynamic balance, alone, did not increase training efficacy. Thus, training for an obstacle avoidance task in a novel visual environment required a critical feature of the criterion task: locomotion. Constant practice with a single lens was successful only if the best lens was selected, but the best lens could not be known ahead of time. Therefore variable practice with multiple lenses on a task that included a critical feature of the criterion task was the best training strategy to enhance adaptive generalization.