A novel manifold learning approach is presented to incorporate computationally efficient obstacle avoidance constraints in optimal control algorithms. The method presented provides a significant computational benefit by reducing the number of constraints required to avoid N obstacles from linear complexity O(N) in traditional obstacle avoidance methods to a constant complexity O(1). The application to autonomous driving problems is demonstrated by incorporation of the manifold constraints into optimal trajectory planning and tracking model predictive control algorithms in the presence of static and dynamic obstacles.