What AIs are not Learning (and Why): Bio-Inspired Foundation Models for Robots

What applications is AI ready for? Advances in deep learning and generative approaches have produced AIs that learn from massive online data and outperform manually built AIs. Some of these AIs outperform people. It is easy (but misleading) to conclude that today's AI technologies are learning to do anything and everything. Conversely, it is striking that big data, deep learning, and generative AI have had so little impact on robotics. For example, today's autonomous robots do not learn to provide home care or to be nursing assistants. Current robot applications are created using manual programming, mathematical models, planning frameworks, and reinforcement learning. These methods do not lead to the leaps in performance and generality seen with deep learning and generative AI. Better approaches to train robots for service applications would greatly expand their social roles and economic impact. AI research is now extending "big data" approaches to train robots by combining multimodal sensing and effector technology from robotics with deep learning technology adapted for embodied systems. These approaches create robotic (or "experiential") foundation models (FMs) for AIs that perceive and act in the world. Robotic FM approaches differ in their expectations, sources, and timing of training data. Like mainstream FM approaches, some robotic FM approaches use vast data to create adult expert-level robots. In contrast, developmental robotic approaches would create progressive FMs that learn continuously and experientially. Aspirationally, these would progress from child-level to student-level, apprentice-level, and expert levels. They would acquire self-developed and socially developed competences. These AIs would model the goals of people around them. Like people, they would learn to coordinate, communicate, and collaborate.