Navigating the Digital World as Humans Do: Universal Visual Grounding for GUI Agents

Multimodal large language models (MLLMs) are transforming the capabilities of graphical user interface (GUI) agents, facilitating their transition from controlled simulations to complex, real-world applications across various platforms. However, the effectiveness of these agents hinges on the robustness of their grounding capability. Current GUI agents predominantly utilize text-based representations such as HTML or accessibility trees, which, despite their utility, often introduce noise, incompleteness, and increased computational overhead. In this paper, we advocate a human-like embodiment for GUI agents that perceive the environment entirely visually and directly perform pixel-level operations on the GUI. The key is visual grounding models that can accurately map diverse referring expressions of GUI elements to their coordinates on the GUI across different platforms. We show that a simple recipe, which includes web-based synthetic data and slight adaptation of the LLaVA architecture, is surprisingly effective for training such visual grounding models. We collect the largest dataset for GUI visual grounding so far, containing 10M GUI elements and their referring expressions over 1.3M screenshots, and use it to train UGround, a strong universal visual grounding model for GUI agents. Empirical results on six benchmarks spanning three categories (grounding, offline agent, and online agent) show that 1) UGround substantially outperforms existing visual grounding models for GUI agents, by up to 20% absolute, and 2) agents with UGround outperform state-of-the-art agents, despite the fact that existing agents use additional text-based input while ours only uses visual perception. These results provide strong support for the feasibility and promise of GUI agents that navigate the digital world as humans do. Install the Township application Web Desktop Turn on Wi-Fi Find the trade-in value for PS4 Mobile Figure 1: Examples of agent tasks across platforms and performance on GUI grounding (♣: ScreenSpot), offline agent (♠: Multimodal-Mind2Web, AndroidControl, and OmniACT), and online agent benchmarks (♥: Mind2Web-Live and AndroidWorld) when using GPT-4 as the planner. Published as a conference paper at ICLR 2025 2. We show that a simple recipe, which includes web-based synthetic data and slight adaptation of the LLaVA architecture (Liu et al., 2024c), is surprisingly effective for GUI visual grounding. Using this recipe, we construct and release the largest GUI visual grounding dataset to date, covering 10M GUI elements and their referring expressions over 1.3M GUI screenshots. We also train and release a universal visual grounding model, UGround, on the dataset. 3. We conduct the most comprehensive evaluation for GUI agents to date, covering six benchmarks spanning three categories (Figure 1 ): grounding (desktop, mobile, and web), offline agent evaluation (desktop, mobile, and web), and online agent evaluation (mobile and web). The results demonstrate: 1) UGround substantially outperforms existing visual grounding models for GUI agents across the board, by up to 20% absolute. 2) SeeAct-V agents with UGround can achieve at least comparable and often much better performance than state-of-the-art agents that use additional text-based input. These results provide strong support for the feasibility and promises of GUI agents that navigate the digital world as humans do. Human-Like Operation Click (556, 26) Type ("4k monitor") User: Decide the next action for the task.