Abstract:To address the insufficient representation of local geometric features in Vision-Language-Action (VLA) models for spatial manipulation tasks, as well as the spatiotemporal mismatch between prediction and execution intervals in real robotic perception-inference-execution pipelines, this paper proposes a lightweight method based primarily on intermediate visual representation alignment, with delay compensation as an auxiliary component. First, a frozen DINOv2 model is adopted as the visual teacher network, and an alignment projection layer is introduced into an intermediate layer of the VLA vision-language backbone. A cosine distance loss is used to constrain the consistency between intermediate visual features and teacher features, thereby enhancing the model’s ability to encode local spatial structures. Meanwhile, a delay compensation mechanism is designed to improve the robustness of the system in real execution pipelines. Experimental results in the LIBERO-Spatial simulation environment show that, after visual representation alignment is introduced, the task success rate of the π0-based model increases from 72.60% to 75.07%, and further increases to 80.87% when combined with delay compensation training. For the π0.5-based model, visual representation alignment improves the task success rate from 94.27% to 96.53%. Real-robot experiments show that, compared with the π0.5 baseline model, the proposed deployment strategy improves the success rates of three spatial manipulation tasks from 70.0%, 55.0%, and 50.0% to 95.0%, 90.0%, and 85.0%, respectively, while reducing the average completion time of successful trials from 17.97s to 16.24s. The results demonstrate that intermediate visual representation alignment can consistently enhance spatial perception across different base models, while delay compensation is more suitable as an auxiliary compensation strategy for real execution pipelines. The proposed method achieves a balance between spatial perception enhancement and real-world deployment adaptability.