Developmental Framework for Autonomous Agent Alignment and Social Norms

The proposed framework addresses the limitations of static datasets in embodied agents by using intrinsic motivations like curiosity to guide environmental interaction. It explores the tension between increasing agent autonomy through direct learning and the increasing difficulty of maintaining human alignment in unstructured environments.