Abstract: Existing six-degree-of-freedom (6-DoF) grasp pose detection methods remain weak in explicitly modeling grasp stability and physical feasibility, making it difficult to ensure high grasp success rates for robotic execution in cluttered scenes. To address this limitation, we propose DPCGrasp, an end-to-end 6-DoF grasp detection method that incorporates differentiable physical constraints. The proposed method introduces four physically motivated constraints: antipodal alignment, surface flatness, center-of-mass proximity, and contact tolerance. These constraints are formulated as differentiable regularization terms and integrated into the training objective to promote physically plausible grasp configurations. To enhance local geometric understanding around candidate grasp points, we design a multi-scale cylindrical sampling and feature fusion module. Furthermore, we develop a self-attention-based multi-parameter grasp prediction head to capture latent dependencies among grasp parameters, improving the consistency of parameter outputs under task-decoupled learning. Experimental results show that the proposed method improves the average precision by 4.66 percentage points on the large-scale GraspNet-1Billion dataset compared to state-of-the-art methods. In real-world robotic experiments, it attains a 7.83 percentage points increase in average grasp success rate, confirming its effectiveness and practical feasibility in actual graspingscenarios.