Aprendizaje por imitación para la locomoción de robots con patas: una revisión

Imitation learning (IL) has fundamentally transformed the field of legged robot locomotion, removing the dependence on hand-engineered reward functions. Since 2019, this area of research has progressed rapidly, from simple motion-capture replication to the generation of sophisticated policies using diffusion models. This survey offers a comprehensive analysis of 35 pivotal research works, using a structured six-dimensional framework to investigate advancements using quadrupedal and humanoid platforms. The review also pinpoints significant challenges related to deployment and outlines new research directions. A key finding from the survey indicates that behavior cloning is utilized in almost half of the analyzed studies. Moreover, data generated through model-predictive control (MPC) now represents the most frequently used training data source for advanced imitation learning systems.