RoboTurk: A Crowdsourcing Platform for Robotic Skill Learning through Imitation

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