In this article we present a new numerical method based on a bilevel decomposition of optimal control problems. A strong connection between the proposed method and the classical indirect multiple shooting method is shown in the regular case, thanks to a link between the Bellman’s value function and the costate from the Pontryagin maximum principle. The value functions are needed in our bilevel decomposition but they are generally difficult to compute. We approximate them by neural networks that have a high potential of generalization and that provide an efficient computation of the gradient of the cost function. We apply the proposed method to an industrial problem, consisting of the determination of torque split and gear shift of a hybrid electric vehicle, the objective being the minimization of the fuel consumption on a given representative cycle. Numerical methods and results are discussed, as well as the possible improvements of the proposed approach.