Actional-Perceptual Causality: Concepts and Inductive Learning for AI and Robotics

Seng-Beng Ho, Mark Edmonds, Song-Chun Zhu

Abstract: Causal learning in AI and robotics has not achieved the same attention as unsupervised learning, supervised learning, and reinforcement learning. The field lacks an agreed upon framework for robotic causal learning. This paper attempts to promote a consensus among researchers in AI and robotics a fundamental framework within which to conduct research on inductive causal learning, i.e., learning causality from data - specifically from perceptual data of an AI or robotic system observing and interacting with the environment. Because actions are a critical aspect of establishing causality for robotic systems, the term actional-perceptual causality is used to emphasize their importance. Using a distinction between diachronic (over time) and synchronic (timeless, contextual) to characterize the causal conditions involved in an actional-perceptual situation, a framework of inductive causal learning for AI and robotics can be made equivalent to the well-established and sound statistical method of causal learning used in experimental sciences, thereby placing actional-perceptual causal learning for AI and robotics on a firm foundation. Actional-perceptual causal learning is a cornerstone of human-like learning and behavior for AI and robotics.