Hierarchies for Embodied Action Perception

摘要: During social interactions, humans are capable of initiating and responding to rich complex actions despite having incomplete world knowledge, physical, perceptual computational constraints. This capability relies on action perception mechanisms that exploit regularities in observed goal-oriented behaviours generate robust predictions reduce the workload sensing systems. To achieve this essential capability, we argue following three factors fundamental. First, human knowledge is frequently hierarchically structured, both execution domains. Second, an active process driven by current task requirements context; particularly important when input (e.g. motion) agent has operate under embodiment Third, learning at heart mechanisms, underlying agent’s ability add new its repertoire. Based these factors, review multiple instantiations a hierarchically-organised biologically-inspired framework for embodied perception, demonstrating flexibility addressing contexts robotic platforms.