Foundational Models for Robotics need to be made Bio-Inspired

Foundation Models for Robotics (FMRs) promise to bring large-scale, generalist intelligence to embodied systems, yet they remain limited in their ability to integrate perception, action, and reasoning in physically grounded environments. In this paper, we argue that advancing FMRs requires drawing inspiration from biological systems - specifically human cognition, development, and sensorimotor learning. We outline five key bio-inspired principles for future FMRs: (1) memory architectures incorporating semantic, episodic, and procedural structures; (2) grounded structured reasoning, as exemplified by embodied chain-of-thought (E-CoT) processes; (3) integration of multimodal sensorimotor feedback, including touch and proprioception; (4) self-motivated learning through simulated play and intrinsic exploration; and (5) neural efficiency through sparse expert activation, functional specialization, and modular reasoning. These elements enable generalization, compositionality, and robustness - traits long demonstrated by humans but underrepresented in current robotic models. While this work does not address reliability and safety in depth, we identify them as essential future directions for developing trustworthy, human-aligned FMRs.