Another promising avenue seeks to ground artificial agents in interactive environments, drawing inspiration from developmental psychology and neuroscience. Embodied AI emphasizes that intelligence arises not from passive observation of data but from active engagement with a physical or simulated world. Agents that learn through sensorimotor interaction, exploration, and goal-directed behavior develop representations tied to causal structures rather than mere correlations. Reinforcement learning in rich environments, such as simulated robotics platforms or procedurally generated worlds, enables the acquisition of skills that transfer more robustly to novel situations. When combined with the predictive objectives of self-supervised learning, these approaches yield internal models that anticipate the consequences of actions, a capacity closely related to planning and intuitive physics.