Infants are highly curious and show remarkable self-driven learning capabilities. Inspired by developmental psychology and recent advances in neuroscience, computational models of curiosity have been implemented in robots. These models are based on the paradigm that learning progress generates intrinsic rewards, which in turn determine action selection. With the rise of deep learning, robots’ perceptual and behavioral learning capabilities have facilitated the appearance of infant-like curiosity-driven behaviors. Implemented in simulation, humanoid robots, drones and aquatic autonomous robots, these curiosity-based models enable open-ended hierarchical learning of skills, which can then be used for extrinsically formulated tasks. In this short review, we highlight the basic components of the most recent curiosity-based models, as well as their implementations in robots that learn about their own body, efficiently map their environment, explore object manipulation and tool use and socially engage with other agents. We conclude with remarks on future directions and challenges.

婴儿非常好奇，并表现出非凡的自我驱动学习能力。受发展心理学和神经科学最新进展的启发，好奇心的计算模型已在机器人中实现。这些模型基于学习进步产生内在奖励的范式，而内在奖励反过来又决定了行动的选择。随着深度学习的兴起，机器人的知觉和行为学习能力促进了婴儿好奇心驱动行为的出现。这些基于好奇心的模型可在仿真，类人机器人，无人机和水上自主机器人中实施，从而使技能的开放式分层学习成为可能，然后可将其用于外部制定的任务。在这篇简短的评论中，我们重点介绍了最新的基于好奇心的模型的基本组成部分，以及它们在机器人中的实现，这些机器人可以了解自己的身体，有效地绘制环境图，探索对象操纵和工具的使用，并与其他代理进行社交互动。最后，我们对未来的方向和挑战发表评论。