A team led by Shen Chunhua at Zhejiang University, in collaboration with the China Telecom Artificial Intelligence Research Institute (TeleAI) and the Gamma Intelligence team led by Bai Chenjia, has proposed a graph reweighting behavior cloning method (GR-BC). The related results have been accepted by CoRL 2026. This method constructs multiple robot trial trajectories for the same task into an association graph containing temporal edges and cross-trajectory similarity edges. It searches in reverse from successful endpoints, computes the minimum path cost from each state to success, and maps it to graph retrieval progress, thereby assigning training weights to effective action segments that satisfy net progress and rollback constraints. This method requires no additional training of reward models or progress predictors and supports the reflow of new trajectories to iteratively update weights. Real-robot experiments cover four types of fine manipulation: four-hole part assembly, graphics card insertion, dual-screw tightening, and memory module insertion. With X-VLA as the backbone, the average success rate is about 91%, significantly outperforming baselines such as DAgger-like methods. It is also compatible with different backbones such as SmolVLA and π₀.₅, can accurately distinguish trajectory success from failure, and capture local regressions during the success process. This method can fully exploit locally effective experience in imperfect demonstrations, providing new ideas for pretraining data cleaning, post-training data quality optimization, and task progress evaluation in robot manipulation, and can support the learning of robust robot policies in industrial-grade millimeter-level fine insertion and assembly scenarios.