Current AI systems lack the 'common sense' understanding of intuitive physics that even young children possess. This paper addresses this gap by drawing inspiration from developmental psychology. A new machine-learning dataset, the Physical Concepts dataset, is introduced to evaluate conceptual understanding of intuitive physics using the violation-of-expectation (VoE) paradigm. A deep-learning system, PLATO, is built to learn intuitive physics directly from visual data, inspired by studies of visual cognition in children. PLATO learns a diverse set of physical concepts, critically depending on object-level representations, mirroring findings from developmental psychology. The implications for both AI and human cognition research are discussed.

Luis S. Piloto, Ari Weinstein, Peter Battaglia, Matthew Botvinick