Scholarlink

<p>The dynamic coupling between motion control and environmental perception remains a critical bottleneck for embodied artificial intelligence (EAI), particularly for systems that operate in unstructured, high-frequency physical environments. Traditional approaches attempt to bridge the “sim-to-real” gap through brute-force domain randomization (DR), implicit system identification, or extensive end-to-end curve fitting. However, these methods predominantly exploit statistical correlations, fundamentally neglecting the underlying physical etiology of dynamic disturbances. Consequently, they suffer from significant performance degradation due to confounding bias when confronted with out-of-distribution (OOD) physical impacts. This paper presents a comprehensive survey that systematically reviews and synthesizes the rapidly expanding literature on the paradigm shift toward causal inference and causal discovery in robotic systems. Drawing upon and consolidating insights from structural causal modeling, interventional reinforcement learning, and physics-grounded control theory, we organize the field into a unified four-pillar taxonomy. We analyze the theoretical limitations of the standard Bellman equation under physical confounding and survey the literature on causal Markov decision processes (C-MDPs). The proposed taxonomy systematically surveys state-of-the-art causal representation learning, utilizing information bottleneck theories and variational inference to extract invariant physical mechanisms from high-dimensional sensor streams. Furthermore, we critically analyze active causal discovery formulated as interventional partially observable Markov decision processes (POMDPs), and evaluate counterfactual reasoning architectures that synthesize OOD recovery behaviors via the abduction-action-prediction calculus. To bridge theory and physical deployment, we extensively review causal-aware closed-loop control, integrating causal control barrier functions (C-CBFs) and introspection in vision-language-action (VLA) models. Finally, we ground these abstractions in detailed hardware case studies—specifically quadrupedal floating-base dynamics and anthropomorphic dexterous manipulation—providing a research blueprint for constructing the next generation of causal-aware embodied foundation models.</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

Beyond correlation: a survey on causal inference for dynamics-aware perception and introspective decision-making in robotics

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • YuLe Cai,
  • Xiang Zhang,
  • Shengrui Liu,
  • Jiacheng Lou,
  • Chunhong Yuan

摘要

The dynamic coupling between motion control and environmental perception remains a critical bottleneck for embodied artificial intelligence (EAI), particularly for systems that operate in unstructured, high-frequency physical environments. Traditional approaches attempt to bridge the “sim-to-real” gap through brute-force domain randomization (DR), implicit system identification, or extensive end-to-end curve fitting. However, these methods predominantly exploit statistical correlations, fundamentally neglecting the underlying physical etiology of dynamic disturbances. Consequently, they suffer from significant performance degradation due to confounding bias when confronted with out-of-distribution (OOD) physical impacts. This paper presents a comprehensive survey that systematically reviews and synthesizes the rapidly expanding literature on the paradigm shift toward causal inference and causal discovery in robotic systems. Drawing upon and consolidating insights from structural causal modeling, interventional reinforcement learning, and physics-grounded control theory, we organize the field into a unified four-pillar taxonomy. We analyze the theoretical limitations of the standard Bellman equation under physical confounding and survey the literature on causal Markov decision processes (C-MDPs). The proposed taxonomy systematically surveys state-of-the-art causal representation learning, utilizing information bottleneck theories and variational inference to extract invariant physical mechanisms from high-dimensional sensor streams. Furthermore, we critically analyze active causal discovery formulated as interventional partially observable Markov decision processes (POMDPs), and evaluate counterfactual reasoning architectures that synthesize OOD recovery behaviors via the abduction-action-prediction calculus. To bridge theory and physical deployment, we extensively review causal-aware closed-loop control, integrating causal control barrier functions (C-CBFs) and introspection in vision-language-action (VLA) models. Finally, we ground these abstractions in detailed hardware case studies—specifically quadrupedal floating-base dynamics and anthropomorphic dexterous manipulation—providing a research blueprint for constructing the next generation of causal-aware embodied foundation models.