RoboTurk: A Crowdsourcing Platform for Robotic Skill Learning through\n Imitation

Abstract

Imitation Learning has empowered recent advances in learning robotic\nmanipulation tasks by addressing shortcomings of Reinforcement Learning such as\nexploration and reward specification. However, research in this area has been\nlimited to modest-sized datasets due to the difficulty of collecting large\nquantities of task demonstrations through existing mechanisms. This work\nintroduces RoboTurk to address this challenge. RoboTurk is a crowdsourcing\nplatform for high quality 6-DoF trajectory based teleoperation through the use\nof widely available mobile devices (e.g. iPhone). We evaluate RoboTurk on three\nmanipulation tasks of varying timescales (15-120s) and observe that our user\ninterface is statistically similar to special purpose hardware such as virtual\nreality controllers in terms of task completion times. Furthermore, we observe\nthat poor network conditions, such as low bandwidth and high delay links, do\nnot substantially affect the remote users' ability to perform task\ndemonstrations successfully on RoboTurk. Lastly, we demonstrate the efficacy of\nRoboTurk through the collection of a pilot dataset; using RoboTurk, we\ncollected 137.5 hours of manipulation data from remote workers, amounting to\nover 2200 successful task demonstrations in 22 hours of total system usage. We\nshow that the data obtained through RoboTurk enables policy learning on\nmulti-step manipulation tasks with sparse rewards and that using larger\nquantities of demonstrations during policy learning provides benefits in terms\nof both learning consistency and final performance. For additional results,\nvideos, and to download our pilot dataset, visit\n$\\href{http://roboturk.stanford.edu/}{\\texttt{roboturk.stanford.edu}}$\n