Scholarlink

Modern autonomous systems are composed of interacting components for control, scheduling, and learning, each operating under practical limitations such as timing uncertainty and imperfect perception. Traditional design and verification approaches aim for component-level perfection, an assumption that is increasingly untenable for complex systems. This paper advocates a shift toward quantitative, system-level safety that explicitly accounts for imperfect components and characterizes how their combined effects impact closed-loop behavior. We introduce safety metrics based on trajectory deviation and reachable set expansion, develop methods to check and synthesize safe schedules under deadline misses, and extend the framework to learning-enabled systems via edge–cloud control and safety-driven resource allocation. Together, these results demonstrate how safe autonomous systems can be systematically designed without requiring their individual components to be perfect.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Building Safe Autonomous Systems Using Imperfect Components

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shengjie Xu,
  • Prateek Ganguli,
  • Tingan Zhu,
  • Arkaprava Gupta,
  • Bineet Ghosh,
  • Kurt Wilson,
  • Abdullah Al Arafat,
  • John Baugh,
  • Zhishan Guo,
  • Benjamin Berg,
  • Parasara Sridhar Duggirala,
  • Samarjit Chakraborty

摘要

Modern autonomous systems are composed of interacting components for control, scheduling, and learning, each operating under practical limitations such as timing uncertainty and imperfect perception. Traditional design and verification approaches aim for component-level perfection, an assumption that is increasingly untenable for complex systems. This paper advocates a shift toward quantitative, system-level safety that explicitly accounts for imperfect components and characterizes how their combined effects impact closed-loop behavior. We introduce safety metrics based on trajectory deviation and reachable set expansion, develop methods to check and synthesize safe schedules under deadline misses, and extend the framework to learning-enabled systems via edge–cloud control and safety-driven resource allocation. Together, these results demonstrate how safe autonomous systems can be systematically designed without requiring their individual components to be perfect.