There is consensus in neuroscience that object understanding arises from hierarchical visual processes, inspiring most of computer vision’s advances in image perception. However, action understanding - how humans perceive and predict the behavior of others - is an embodied process. Mounting evidence from cognitive science shows that observing others’ actions also engages our motor systems, suggesting that understanding action involves both seeing and simulating movement. We aim to create a framework inspired by the rich interplay between perceptual processes and motoric processes in the brain, based on two prior works of ours - LEAP and EmbodiSwap. LEAP uses LLMs to translate egocentric human videos into structured, symbolic action programs. EmbodiSwap composites robot embodiments into human videos for zero-shot robot imitation learning. Combining these works, we propose a framework in which visual processes that extract programs of action from human video benefit from embodied interaction enabled through the deployment of visuomotor policies. Through this framework, we aim to demonstrate that a real-world robot can actively refine its visual representations through action.