USC researcher demos teleop stack
USC assistant professor Yue Wang shared a demo and notes on scaling humanoid pre‑training using egocentric video plus a single‑operator teleoperation rig that pairs a PICO headset, MANUS gloves, and trackers for full‑body dexterous control. The post includes a video showing manipulation-focused teleop data collection for humanoid pretraining. (x.com)
Yue Wang runs the Physical SuperIntelligence Lab at USC and lists both USC affiliation and prior NVIDIA Research connections on his lab page and personal site. (yuewang.xyz) A related preprint titled “Learning from Massive Human Videos for Universal Humanoid Pose Control” (arXiv:2412.14172) formalizes the paper’s methodology of combining large-scale egocentric video with teleoperation data for scalable humanoid pretraining and was posted to arXiv in December 2024. (arxiv.org) USC’s Humanoid-Teleop repository publicly documents a custom data-collection pipeline and explicitly states it includes re‑engineering work for Unitree’s G1/H1 teleoperation branches, indicating the demo ties into an open codebase for humanoid teleop. (github.com) NVlabs’ GR00T and related Pico controller documentation show established patterns for using PICO headsets to stream whole-body SMPL poses for humanoid teleoperation, matching the headset choice implied in the demo stack. (nvlabs.github.io) MANUS markets high‑precision data gloves aimed at robotics and teleoperation use cases with low-latency finger pose streams, and MANUS integration guides are commonly referenced in XR-to-robot teleop workflows. (manus-meta.com) A Unitree/xr_teleoperate issue thread and several community repos show active developer work to adapt Quest/PICO flows for teleoperation, underscoring that single‑operator XR+glove+tracker setups are an emerging, community-supported approach for manipulation data collection. (github.com)
