The broader goal of our research is to train infants with special needs to safely and purposefully drive a mobile robot to explore the environment. The hypothesis is that these impaired infants will benefit from mobility in their early years and attain childhood milestones, similar to their healthy peers. In this paper, we present an algorithm and training method using a force-feedback joystick with an “assist-as-needed” paradigm for driving training. In this “assist-as-needed” approach, if the child steers the joystick outside a force tunnel centered on the desired direction, the driver experiences a bias force on the hand. We show results with a group study on typically developing toddlers that such a haptic guidance algorithm is superior to training with a conventional joystick. We also provide a case study on two special needs children, under three years old, who learn to make sharp turns during driving, when trained over a five-day period with the force-feedback joystick using the algorithm.