Hybrid wheeled-legged robots offer superior mobility in unstructured terrain by combining the locomotion efficiency of wheels with the adaptability of legs. However, efficiently traversing complex obstacles necessitates intelligent reconfiguration policies that coordinate leg movements, wheel velocities, and body posture. This paper presents an integrated learning framework to derive optimal policies for a six-wheeled-legged robot traversing an asymmetric step obstacle. Initially, the kinematic relationships governing wheel-leg interaction with the ground and slip-free wheel velocity coordination are formulated. As the key innovation, the obstacle traversal problem is mapped into a low-dimensional and continuous parametric optimization space, upon which a Genetic Algorithm-based learning framework is developed. Without imposing restrictive assumptions, this framework significantly reduces the complexity of the search space. In addition, fuzzy logic is employed to integrate human expertise into the cost function and adapt its weighting coefficients. This facilitates the automatic discovery of practical policies without relying on pre-defined motion sequences. Simulation results are validated through experiments on the ViraHex robot, demonstrating close agreement in leg motion sequences, support polygon transitions, and the mitigation of undesired body deviations. Overall, this framework offers an effective and scalable solution for optimizing obstacle traversal in hybrid robots, highlighting the potential of learning-based strategies in complex environments.