The Future of Research in Cognitive Robotics: Foundation Models or Developmental Cognitive Models?

Robotics research has recently embraced foundation models as an attractive way to enhance different elements of the robotics stack and to simplify the challenge of instructing robots to perform tasks. While there is increasing evidence that this may be a very useful approach, there are also many concerns with making this the exclusive route to the future of robotics, e.g., cost, availability, trustworthiness, robustness, transparency, security, and inclusion. In this article, we will explore an alternative future for robotics research, one which eschews foundation models and deep learning with internet‐scale datasets, and adopts instead an approach that has more in common with cognitive development in humans, exploiting knowledge of cognitive science while at the same time using experience in building cognitive robots to make contributions to both cognitive robotics and cognitive science. Such approaches are more aligned with the enactive stance in embodied cognitive science, in contrast with the disembodied computational functionalism of cognitivist approaches, of which foundation models are the most contemporary instance.