Autonomous agents capable of operating real-world graphical user interfaces (GUIs) depend critically on accurate perception of on-screen elements. While recent vision-language models (VLMs) exhibit strong reasoning capabilities, unreliable detection and localization of interactive components remain a key limitation. Existing systems such as OmniParser rely on anchor-based detection and non-maximum suppression (NMS), which degrade spatial precision and consistency in dense, overlapping, or low-contrast layouts. In this paper, we present CU-1, a class-agnostic Detection Transformer (DETR) designed for autonomous UI agents. By treating all interface elements as instances of a single object class and focusing model capacity on spatial precision rather than categorical discrimination, CU-1 achieves robust localization across heterogeneous UI designs. On the WebClick benchmark, CU-1 attains 70. 8% end-to-end task success, a 12 percentage point improvement over OmniParser V2 (58. 8 \%), representing a 20. 4 % relative gain. In dense calendar interfaces containing over 80 small elements per screen, CU-1 reaches 64% success versus 46% for OmniParser V2; a 39% relative improvement attributable to transformer attention mechanisms. Training converges within 90 minutes on a single H100 GPU (50 epochs, 0. 42 mAP @ 0. 5), and real-time inference under 100 ms enables fluid agent-environment interaction on standard hardware. CU-1 establishes a practical and efficient foundation for scalable multimodal agents by aligning visual perception with the end-to-end requirements of natural-language-driven UI automation. The model is available at: https: //huggingface. co/racineai/CU-1.
Rochet et al. (Mon,) studied this question.