Automatic free-form surface reconstruction enables flexible and efficient industrial manufacturing. However, due to the performance of the 3-D measurement system, accurate and robust reconstruction is quite difficult due to corrupted data, discontinuous scanning, etc. To address these issues, in this article, we propose a novel gravitational discriminative optimization (GDO) method based on a multiview reconstruction framework for free-form blades. Our method consists of a training phase and a reconstruction phase. In the training phase, we introduce a novel gravitational distribution feature that is robust against the perturbations caused by corrupted data to instruct the sequence of update maps (SUM) learning process of GDO. To deal with the low overlap caused by the limited and discontinuous scanning, we propose a sequential scenes-to-model framework. Equipped with the trained SUM, we design the reconstruction phase to perform the blade model reconstruction accurately and robustly by partitioning it into inference with a gravitation map, an adaptive maps scheme to boost convergence, and an iterative closest point refinement. Experiments on both limited scanning-based synthetic data and real scene reconstruction of four blade models with different profiles show that GDO outperforms five other methods in terms of both accuracy and robustness.