End-to-end autonomous driving has witnessed significant advancements in recent years. In this work, we present OAD, an enhanced framework built upon the VAD architecture, which introduces a novel paradigm shift from direct trajectory prediction to trajectory offset learning. This approach significantly improves planning accuracy and safety, reducing the average planning displacement error by 39.7% (from 0.78m to 0.47m) and the collision rate by 84.2% (from 0.38% to 0.06%) compared to the original VAD model. Furthermore, OAD demonstrates the capability to generate diverse, multi-modal trajectories, enhancing its adaptability to complex driving scenarios.