Haptic guidance can improve the immediate performance of a motor task by enforcing a desired pattern of kinematics, but several studies have found that it impairs motor learning. In this study, we studied whether guidance from a robotic steering wheel can improve one's short-term learning of steering a simulated vehicle. We developed a computer algorithm that adapted the firmness of the guidance based on ongoing error. Training with "guidance-as-needed" or "fixed guidance" allowed participants to learn to steer without experiencing large errors and produced slightly better immediate retention than did training without guidance, apparently because participants were better able to learn when to initiate turns. Training with guidance-as-needed produced slightly better results than training with fixed guidance: the guidance-as-needed participants' errors were significantly smaller when guidance was removed. However, this difference quickly dissipated with more practice. We conclude that haptic guidance can benefit short-term learning of a steering-type task while also limiting performance errors during training.