To address the challenges of data scarcity and simulation-to-reality domain shift in robotic systems operating in complex environments, this paper proposes a cross-domain recognition framework integrating dynamic domain randomization (DDR) and domain adversarial transfer learning (DATL). A multimodal perception system incorporating visual, force, and proximity sensing is established, complemented by a fusion network enabling cross-modal information complementarity. By incorporating gradient reversal layers (GRL), we construct a domain adversarial transfer architecture to mitigate feature distribution discrepancies between simulated and real-world scenarios. Experimental results demonstrate that the proposed method achieves 92.32% object classification accuracy in real-world scenarios, representing improvements of 52.67% and 55.92% compared to ResNet-50 with visual-only data and ResNet-50 with multimodal data baselines, respectively. These results validate the effectiveness of our integrated DDR and DATL approach in bridging the sim-to-real gap in robotic perception.