Molecular docking, which predicts the bound structure of protein-ligand conformations, is essential for understanding cellular mechanisms and plays a crucial role in applications such as structure-based drug design. Recent advances in generative modeling, such as flow matching, have achieved simple and effective solutions for this task by modeling docking conformations as a distribution. In this work, we focus on flow matching generative models and adopt an energy-based perspective for understanding the confidence model. A mapping function, represented by a deep network, is directly learned to iteratively map random configurations, i.e., samples from the source distribution, to bound structures, i.e., points in the target data manifold. This yields a conceptually simple and empirically effective flow matching setup with interesting connections to fundamental notions in generative modeling such as idempotency and stability as well as empirical structure prediction techniques such as refinement. Experiments on PDBBind and Binding MOAD for both single and multi-ligand docking consistently demonstrate the method's effectiveness where it outperforms recent baselines of standard flow matching and task-associated diffusion model, using similar computational budget.