MIMo virtual infant learns to roll over via reinforcement learning

A team led by Leon Philipp, Francisco M. López, and Jochen Triesch at Goethe University Frankfurt has created MIMo, a multimodal virtual infant model that learns to roll over from supine to prone using reinforcement learning. MIMo is equipped with realistic proprioception (sense of limb position) and vestibular sensation (balance and orientation). The model’s learned rolling behaviors capture key developmental trends observed in real infants: execution speed improves over simulated age, and coordination patterns become more efficient. The study, accepted at the 2026 IEEE ICDL Conference, emphasizes that the infant’s changing body morphology—growth in limb length, mass, and joint stiffness—directly shapes the emergence of rolling as a whole-body sensorimotor skill.

This work demonstrates how embodied computational models can bridge robotics and developmental psychology. By constraining the algorithm with physically realistic body changes, the team achieved behaviors that are not just simulated but biologically plausible. The findings suggest that AI systems may benefit from similar developmental constraints when learning motor tasks. MIMo could serve as a testbed for studying early motor milestones without the ethical and practical challenges of real infant experiments, opening new avenues for understanding human development and for designing more adaptable robots.