DesignLex
/人机交互/Computer Use Agent (CUA)
论文进阶

Computer Use Agent (CUA)

计算机使用代理(CUA):通过解读屏幕截图与界面结构(如 DOM、可访问性树),并执行点击、输入、滚动等底层操作来自动完成用户任务的 AI 代理。

English Definition

"An AI agent that interacts with computer applications by interpreting screen pixels and structural context (DOM, accessibility trees), then performing low-level operations such as clicking, typing, and scrolling to autonomously complete user tasks."

用法说明 · 针对中文母语者

伴随多模态大模型(如 Claude、GPT-4V)发展而来的新兴 HCI/AI 术语。代表作包括 OpenAI Operator、Anthropic Computer Use、Microsoft Copilot、Google Project Astra 等。注意与 RPA(机器人流程自动化)的区别——CUA 基于视觉理解和 LLM 推理,而非预先脚本。

常见错误

  • ·误用为通用'AI 助手'——CUA 特指能像人一样操作 GUI 的代理
  • ·与 chatbot 混淆——CUA 强调对界面执行操作,而非仅生成文本

真实用例 · 4

"Computer Use Agents (CUAs) operate interfaces by pointing, clicking, and typing—mirroring interactions of sighted users (SUs)..."

论文Pipeline extracted — review needed·给出 CUA 的核心定义和操作方式

"CUAs such as OpenAI's Operator, Anthropic's Computer Use Tool, Microsoft Copilot, and Google DeepMind's Project Astra have demonstrated this emerging capability."

论文Pipeline extracted — review needed·列举当前代表性的 CUA 产品

"Current CUAs however, overlook interaction practices of blind and low-vision users (BLVUs) who use assistive technologies (AT)..."

论文Pipeline extracted — review needed·指出当前 CUA 在无障碍场景下的局限性

"An evaluation of state-of-the-art CUAs under default and AT (keyboard-only, magnified viewport) conditions, revealing accessibility gaps in current CUAs."

论文Pipeline extracted — review needed·在贡献列表中说明对 CUA 的评估是论文核心贡献之一
由 pipeline 自动采集,待人工 review