Mosaic: Scalable Robotics Data Collection Platform
Author(s)
Gandhi, Ansh
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
The lack of large, high-quality human demonstration datasets remains a key obstacle to advancing imitation learning for robotic manipulation. Crowdsourcing data through accessible devices like smartphones offers a promising solution. This work presents a scalable data collection platform that enables cloud-based teleoperation of simulated robots using GPU-accelerated infrastructure and diverse input modalities, including dual-smartphone input for bimanual control. By integrating a user training curriculum, systematic performance metrics, and comprehensive user studies, the platform significantly lowers the barrier to contributing high-quality demonstrations, advancing scalable data collection for imitation learning and beyond.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis