Many real-world decisions are made under uncertainty by solving optimization problems using predicted quantities. This predict-then-optimize paradigm has motivated decision-focused learning, which trains models with awareness of how the optimizer uses predictions, improving the performance of downstream decisions. Despite its promise, scaling is challenging: state-of-the-art methods either differentiate through a solver or rely on task-specific surrogates, both of which require frequent and expensive calls to an optimizer, often a combinatorial one. In this paper, we leverage dual variables from the downstream problem to shape learning and introduce Dual-Guided Loss (DGL), a simple, scalable objective that preserves decision alignment while reducing solver dependence. We construct DGL specifically for combinatorial selection problems with natural one-of-many constraints, such as matching, knapsack, and shortest path. Our approach (a) decouples optimization from gradient updates by solving the downstream problem only periodically; (b) between refreshes, trains on dual-adjusted targets using simple differentiable surrogate losses; and (c) as refreshes become less frequent, drives training cost toward standard supervised learning while retaining strong decision alignment. We prove that DGL has asymptotically diminishing decision regret, analyze runtime complexity, and show on two problem classes that DGL matches or exceeds state-of-the-art DFL methods while using far fewer solver calls and substantially less training time. Code is available at https://github.com/

Aiming at the traditional grasping method for manipulators based on 2D camera, when faced with the scene of gathering or covering, it can hardly perform well in unstructured scenes that appear as gathering and covering, for the reason that can't recognize objects accurately in cluster scenes from a single perspective and the manipulators can't make the environment better for grasping. In this case, a novel method of pushing-grasping collaborative based on the deep Q-network in dual perspectives is proposed in this paper. This method adopts an improved deep Q network algorithm, with an RGB-D camera to obtain the information of objects' RGB images and point clouds from two perspectives, and combines the pushing and grasping actions so that the trained manipulator can make the scenes better for grasping so that it can perform well in more complicated grasping scenes. What's more, we improved the reward function of the deep Q-network and propose the piecewise reward function to speed up the convergence of the deep Q-network. We trained different models and tried different methods in the V-REP simulation environment, and it concluded that the method proposed in this paper converges quickly and the success rate of grasping objects in unstructured scenes raises up to 83.5%. Besides, it shows the generalization ability and well performance when novel objects appear in the scenes that the manipulator has never grasped before.