RoboTurk: A Crowdsourcing Platform for Robotic Skill Learning through Imitation (1811.02790v1)

Published 7 Nov 2018 in cs.RO, cs.AI, and cs.LG

Abstract: Imitation Learning has empowered recent advances in learning robotic manipulation tasks by addressing shortcomings of Reinforcement Learning such as exploration and reward specification. However, research in this area has been limited to modest-sized datasets due to the difficulty of collecting large quantities of task demonstrations through existing mechanisms. This work introduces RoboTurk to address this challenge. RoboTurk is a crowdsourcing platform for high quality 6-DoF trajectory based teleoperation through the use of widely available mobile devices (e.g. iPhone). We evaluate RoboTurk on three manipulation tasks of varying timescales (15-120s) and observe that our user interface is statistically similar to special purpose hardware such as virtual reality controllers in terms of task completion times. Furthermore, we observe that poor network conditions, such as low bandwidth and high delay links, do not substantially affect the remote users' ability to perform task demonstrations successfully on RoboTurk. Lastly, we demonstrate the efficacy of RoboTurk through the collection of a pilot dataset; using RoboTurk, we collected 137.5 hours of manipulation data from remote workers, amounting to over 2200 successful task demonstrations in 22 hours of total system usage. We show that the data obtained through RoboTurk enables policy learning on multi-step manipulation tasks with sparse rewards and that using larger quantities of demonstrations during policy learning provides benefits in terms of both learning consistency and final performance. For additional results, videos, and to download our pilot dataset, visit $\href{http://roboturk.stanford.edu/}{\texttt{roboturk.stanford.edu}}$

Authors (12)

Ajay Mandlekar (41 papers)
Yuke Zhu (134 papers)
Animesh Garg (129 papers)
Jonathan Booher (5 papers)
Max Spero (5 papers)
Albert Tung (8 papers)
Julian Gao (2 papers)
John Emmons (6 papers)
Anchit Gupta (21 papers)
Emre Orbay (1 paper)
Silvio Savarese (200 papers)
Li Fei-Fei (199 papers)

Citations (254)

View on Semantic Scholar

Summary

An Expert Overview of RoboTurk: Expanding Data Collection for Imitation Learning in Robotics

In the domain of robotic manipulation, Imitation Learning (IL) has been highlighted as a potent alternative to Reinforcement Learning (RL), particularly where environment interactions are costly and reward specifications pose significant challenges. Researchers Mandlekar et al. introduce "RoboTurk", a scalable crowdsourcing platform devised to alleviate the impediments linked with collecting vast amounts of demonstration data for IL, particularly within the arena of high degree-of-freedom problem spaces. Leveraging ubiquitous mobile technology, such as smartphones, RoboTurk facilitates the collection of quality teleoperated demonstrations necessary for training robust IL models.

The paper systematically evaluates RoboTurk’s effectiveness across an array of manipulation tasks, meticulously examining both interface usability and resilience under diverse network conditions. Remarkably, the researchers demonstrate that smartphone-driven teleoperation yields performance metrics statistically indistinguishable from specialized VR controllers, thereby democratizing data collection without sacrificing task execution quality. The resilience of RoboTurk is underscored through testing under high-latency and low-bandwidth conditions, where performance metrics remain comparably robust.

RoboTurk's integration of cloud-based simulation and WebRTC protocols ensures an expansive reach to various users, diminishing hardware barriers traditionally limiting widespread crowdsourcing adoption. The platform’s architecture allows real-time manipulation of simulated robotic systems, facilitated through an intuitive user interface that exploits the existing user hardware ecosystem. At a design level, RoboTurk’s infrastructure—including its coordination server and teleoperation server—ensures scalability without compromising on latency, making it adaptable for a multitude of robotic tasks and simulations.

The authors substantiate their claims through empirical data derived from a pilot dataset encompassing 137.5 hours of demonstration trajectories, procured within merely 22 hours of total system usage. This dataset enabled evaluation of policy learning, demonstrating that increased quantities of demonstrations significantly augment learning consistency and the converged performance of robotic policies, especially under sparse reward conditions. Notably, experiments on bin picking and nut-and-peg assembly tasks reveal that policy performance scales with the volume of demonstration data utilized.

From a theoretical standpoint, RoboTurk represents a paradigm shift toward gathering large-scale, high-fidelity datasets that could reshape how policy learning is approached. By providing scalable IL data, RoboTurk reduces the dependence on extensive RL environment interactions, offering a practical bridge between demonstration-driven learning and complex behavior modeling. Practically, this platform paves the way for applying learned skills in variable, real-world scenarios where adaptability and precision are paramount.

Looking forward, the implications of this research herald new avenues in improving IL algorithms, especially as RoboTurk’s inherently modular design encourages further extension to physical robotic interactions. There remains potential for integrating more expansive and diverse task sets, which could enhance the generalization capabilities of IL models across different domains. Furthermore, optimizing the quality of data collection by utilizing enhanced methodologies in augmented reality and haptic feedback could refine the fidelity of teleoperated robot control.

In summary, the introduction of RoboTurk provides an accessible, robust platform for scaling demonstration data collection in robotic IL, setting a foundational premise for future explorations in data-driven robotics. With its proven utility and implications for improved model training efficiency, RoboTurk could transform the landscape of robotic learning and manipulation capabilities.

PDF Markdown

Related Papers

Find Related Papers