In this paper，a single-stage domain adaptive commodity recognition algorithm YOLO-DA（YOLO-domain adaptive）is proposed. Firstly，adaptation adjustments are made to the YOLO algorithm for cross-domain tasks and the RPC dataset. Secondly，the neck network structure is redesigned，incorporating the BiFPN concept to re-fuse features at multiple scales. Finally，a Gradient Reversal Layer is added behind the backbone network for adversarial training on the training and testing sets，further approaching the goal of domain adaptation. The training results of the improved network model on the RPC dataset show that the mean average precision（mAP）reaches 65.25%. Compared with the baseline network，the detection accuracy is significantly improved，and the cases of missed detection and false detection are notably reduced.