Scholarlink

<p>Non-stationary physical environments—arising from changes in friction, payload, contact conditions, human intervention, and system degradation—remain a central obstacle to deploying learning-enabled controllers in safety–critical settings. While substantial progress has been made in learning-based control and reinforcement learning, many failures under distribution shift can be traced back to a modeling gap: the inability to infer and track the <i>physical causes</i> that deform system dynamics and safety constraints over time. This review consolidates recent advances in <i>physics-informed causal reasoning and modeling</i> for non-stationary environments, with an emphasis on (<i>i</i>) inferring safety-relevant physical context from observations and interaction (action–reaction), (<i>ii</i>) building context-modulated dynamics models with physics priors injected via losses, architectures, or explicit constraints, and (<i>iii</i>) quantifying and calibrating uncertainty in <i>actionable</i> forms that support downstream safety mechanisms, including constraint tightening, risk-sensitive formulations, and distributionally robust reasoning. We propose a unified taxonomy that disentangles <i>task-relevant state</i> from <i>safety–critical physical context</i> and highlights failure modes induced by appearance–physics mismatch and non-informative cues. We organize the literature along axes of non-stationarity types, context identification paradigms (passive vs. active), and physics-prior injection mechanisms. Beyond summarizing methods, we provide an evaluation checklist and reporting protocol to improve comparability across studies, and conclude with open problems and a research agenda toward deployment-ready <i>physical AI systems grounded in real-world physics</i>.</p>

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

Physics-Informed Causal Reasoning in Physical AI: A Review on Modeling Non-stationary Environments for Safety–Critical Control

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Sung Kwon On,
  • Sukchul Jeong,
  • Jungjin Lee,
  • Insoon Yang

摘要

Non-stationary physical environments—arising from changes in friction, payload, contact conditions, human intervention, and system degradation—remain a central obstacle to deploying learning-enabled controllers in safety–critical settings. While substantial progress has been made in learning-based control and reinforcement learning, many failures under distribution shift can be traced back to a modeling gap: the inability to infer and track the physical causes that deform system dynamics and safety constraints over time. This review consolidates recent advances in physics-informed causal reasoning and modeling for non-stationary environments, with an emphasis on (i) inferring safety-relevant physical context from observations and interaction (action–reaction), (ii) building context-modulated dynamics models with physics priors injected via losses, architectures, or explicit constraints, and (iii) quantifying and calibrating uncertainty in actionable forms that support downstream safety mechanisms, including constraint tightening, risk-sensitive formulations, and distributionally robust reasoning. We propose a unified taxonomy that disentangles task-relevant state from safety–critical physical context and highlights failure modes induced by appearance–physics mismatch and non-informative cues. We organize the literature along axes of non-stationarity types, context identification paradigms (passive vs. active), and physics-prior injection mechanisms. Beyond summarizing methods, we provide an evaluation checklist and reporting protocol to improve comparability across studies, and conclude with open problems and a research agenda toward deployment-ready physical AI systems grounded in real-world physics.