RoboTurk

Stanford's 2018 platform for crowdsourced robot demonstrations, teleoperated remotely from a smartphone.

RoboTurk is a crowdsourced data-collection platform published at CoRL 2018 by Fei-Fei Li, Silvio Savarese, and colleagues at Stanford, named after the crowdsourcing platform Amazon Mechanical Turk. It lets ordinary people perform 6-DoF teleoperation using a smartphone like an iPhone: moving the phone through the air makes the remote robot's end effector move the same way, with no VR headset or special equipment needed, so collection tasks can be distributed to remote workers anywhere. In an initial trial totaling 22 hours of system usage, it collected 137.5 hours of operation data and over 2,200 successful demonstrations; the paper also found that remote users could complete demonstrations successfully even with low bandwidth and high latency. It's a representative example of crowdsourced collection, and the multi-operator demonstration data in the robomimic benchmark was also collected with it.

ExampleA remote worker opens the app at home, watching a live video feed streamed back over the web while moving their phone to control a robot arm through a pick-and-place task; the data uploads automatically for imitation learning.

Also called

RoboTurk Crowdsourcing Platform

Related

Crowdsourced Data Collection · Teleoperation · Demonstration Data · robomimic · Imitation Learning · Stanford Artificial Intelligence Laboratory

Sources

RoboTurk: A Crowdsourcing Platform for Robotic Skill Learning through Imitation (arXiv:1811.02790)

As of

2018-11