
Modeling Humans for Humanoid Robots
Xiaolong Wang of UC San Diego presents a two-level learning framework for giving humanoid robots robust mobility and manipulation skills, part of Stanford's AA289 Robotics and Autonomous Systems Seminar. The first level trains Vision-Language-Action models on human video data, predicting mid-level body and hand trajectories conditioned on language instructions, for both navigation and manipulation tasks. The second level builds low-level skills through teleoperation and whole-body humanoid control via motion imitation and Sim2Real transfer. Wang explains how combining human-derived VLA models with these low-level robot skills offers a scalable path toward general-purpose humanoid robots that can generalize across tasks, objects, and environments. The talk covers the technical pipeline from human demonstration data to deployed robot behavior, situating the work within the long-standing robotics goal of human-like operation in diverse daily-life settings.