Title: Scalable Motor Skill Learning for Diverse Robot Embodiments

Abstract: Inspired by advancements in large-scale models for vision and language, recent efforts have aimed to scale imitation learning by collecting demonstrations across diverse embodiments and tasks. However, unlike vision and language domains, robotics requires physical interaction with the environment, which introduces significant complexities to motor skill learning and incurs substantial costs associated with robot hardware and human workload for large-scale data acquisition. To address these challenges, my research focuses on developing cost-efficient learning frameworks that minimize data collection costs while enabling the reusability of learned motor skills across various robot embodiments. By leveraging robust controllers capable of translating high-level commands into complex morphologies, skills can be abstracted, trained in cost-effective domains, and transferred to different robotic platforms. Furthermore, intuitive data-collection interfaces facilitate scalable demonstration collection while reducing human workload. In this talk, I will present three of my recent works aimed at cost-effective motor skill acquisition: 1) a hybrid learning algorithm that leverages the complementary strengths of simulation and real-world data collection [link], 2) a skill-learning framework enabling complex humanoid loco-manipulation through whole-body control [link], and 3) a cross-embodiment learning framework utilizing robust imitation learning for transferable visuomotor skills, combined with IK-based whole-body motion retargeting [link].

Bio: Mingyo Seo is a PhD student at the University of Texas at Austin, advised by Yuke Zhu and Luis Sentis. His research focuses on developing skill-learning frameworks for diverse and complex robotic embodiments. His research topics include robot learning, control, and motion planning. For more information, visit his website: https://mingyoseo.com.