Causal artificial intelligence for robust robot reasoning under uncertainty

Thesis

Autonomous robots must operate effectively in complex, uncertain, and partially observable environments. Achieving this requires the ability not only to perceive and act, but also to understand and reason about cause and effect. This thesis investigates how causal generative AI can provide a unified foundation for robust robot cognition under uncertainty. It introduces a conceptual framework for Robot Causal Reasoning that integrates structural causal models, probabilistic planning, and (deep and classical) generative world modelling to enhance robot perception, inference, decision-making, and explanation. Across six research questions, the work spans interventional and counterfactual reasoning, advancing through causal extensions to planning, manipulation, and generative modelling. The thesis contributes new methods at each level of Pearl’s Ladder of Causation. At the interventional level, for the first time we extend online POMDP planning to reason about confounded dynamics, and introduce a causal Bayesian reasoning architecture for manipulation under uncertainty. At the counterfactual level, the thesis develops methodology for post-hoc causal robot explanations and extends causal reasoning into high-dimensional deep generative AI through counterfactual contrastive learning and a parallel-world simulation framework. Together, these advances establish a coherent causal hierarchy unifying modelling, learning, inference, and explanation. They demonstrate that causal generative models can bridge symbolic reasoning and data-driven learning, enabling robots to act, imagine, and explain in ways consistent with human causal understanding. By grounding autonomous behaviour in causal structure, this work contributes to the next generation of trustworthy cognitive robotics: systems that are not only capable of robust decision-making, but also of understanding and communicating the reasons for their actions.

Authors: Cannizzaro, R

• DOI: 10.5287/ora-azok5kr7g
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: modelling; reasoning; cognitive robotics; causal attribution; artificial intelligence; counterfactual reasoning; autonomous systems; causality; probabilistic programming; representation learning; Bayesian inference; contrastive learning; agentic AI; explainable AI; generative AI; latent diffusion; robotics; Pearl's ladder of causation; world models; causal machine learning; video game world models; simulation; causal models; deep learning