跳转至

运行级数字孪生诊所：实现具身智能基于任务的评估

文章背景与核心概要

在真实临床环境中直接测试具身人工智能（AI）至关重要，但构建真实且适合机器人测试的医疗场景往往成本高昂且难以规模化。为此，本研究引入了一种创新方法，可将常规的诊所照片转化为可进行任务评估的运行级数字孪生体。

研究团队利用 39 个真实的眼科诊所场景，成功将单张图像捕获转换为可编辑、可直接用于模拟器的环境。其评估框架全面涵盖了重建质量、房间级几何形状、网格锚定（mesh grounding）、多机器人可行性、扰动敏感度以及闭环策略性能。该工作确立了“运行有效性”作为临床数字孪生核心设计原则的重要地位，有效架起了具身智能离线开发与物理医疗部署之间的桥梁。

Testing embodied artificial intelligence (AI) directly in real-world clinical environments is essential, yet creating authentic, robot-testable healthcare settings remains expensive and difficult to scale. This research introduces a methodology to transform routine clinic photographs into operational digital twins for task-based evaluation of embodied AI.

Utilizing 39 actual ophthalmic clinic scenes, the authors successfully converted single-image captures into editable, simulator-ready environments. Their evaluation framework assesses reconstruction quality, room-scale geometry, mesh grounding, multi-robot feasibility, perturbation sensitivity, and closed-loop policy performance.

arXiv ID: [arXiv:2608.21416 [cs.RO]]
Primary Subject: Robotics (cs.RO)
Secondary Subject: Artificial Intelligence (cs.AI)
Submission Date: August 12, 2026
Authors: Xinyuan Wu, Jingrao Zhang, Mengdi Xu, Henry K. Chu, Mingguang He, Danli Shi

arXiv ID: [arXiv:2608.21416 [cs.RO]]
Primary Subject: Robotics (cs.RO)
Secondary Subject: Artificial Intelligence (cs.AI)
Submission Date: August 12, 2026
Authors: Xinyuan Wu, Jingrao Zhang, Mengdi Xu, Henry K. Chu, Mingguang He, Danli Shi

Summary

关键发现：

  • 保留的工作空间结构： 重建后的场景精确保持了工作空间的几何形状，同时允许通过可编辑的局部修改功能进行本地设备重配置。
  • 具备接触感知能力的仿真： 设备网格、碰撞代理（collision proxies）和语义锚点的集成，使视觉重建能够无缝转化为具有物理交互能力的环境。
  • 具身特异性洞察： 在三种不同的机器人形态下，共享的任务目标揭示了可达性和接触可行性的独特模式。
  • 超越视觉相似性的敏感度： 设备的微小平移和旋转会产生特定于任务的接触边界变化，这是仅凭视觉相似性无法检测到的。
  • 策略学习： 数字孪生轨迹成功支持了局部策略学习和闭环性能评估。

最终，这项工作确立了运行有效性（operational validity）作为临床数字孪生的基础设计原则，弥合了具身智能离线开发与物理医疗部署之间的鸿沟。

Key Findings:

  • Preserved Workspace Structure: The reconstructed scenes accurately maintain workspace geometry while allowing local device reconfigurations via editable local editing features.
  • Contact-Aware Simulation: Integration of device meshes, collision proxies, and semantic anchors seamlessly translates visual reconstructions into physically interactive environments.
  • Embodiment-Specific Insights: Across three distinct robot embodiments, shared task targets revealed unique patterns of reachability and contact feasibility.
  • Sensitivity Beyond Visual Similarity: Minor translations and rotations of devices produced task-specific alterations in contact margins that could not be detected through visual similarity alone.
  • Policy Learning: Digital-twin trajectories successfully supported local policy learning and closed-loop performance evaluation.

Ultimately, this work establishes operational validity as a foundational design principle for clinical digital twins, bridging the gap between offline development and physical healthcare deployment for embodied AI.

Article Metadata & Links

文章元数据与链接

External References & Resources

外部参考与资源