Machine learning has yet to transform robotics the way it has transformed computer vision. The startup costs to creating a new robotics lab are prohibitively high: new investigators must invest not only hundreds of thousands of dollars for new robots and lab space, but also years of cumulative research hours setting up a new robot, calibrating it, and reimplementing baselines just to begin contributing to the research frontier. Similarly, there is little guarantee that an algorithm achieving impressive results in one lab’s setting will work well in others’, due to myriad differences in robotic hardware, visual and physical properties of the environment, and other implementation details. As a result, even potential breakthroughs struggle to gain broader traction. In practice, this introduces a rich gets richer bias: big and established labs can focus on optimizing methods for their own setup while comparing against their own past work. In contrast, new entrants find it needlessly hard to reproduce simple experiments, let alone push the state of the art. It is clear that the status quo needs to change.


Many roboticists have documented these challenges and have over the years tried to address them through standardized hardware (YCB, PR2, Sawyer), standardized software (ROS, PyRobot), self-supervised data collection (Arm Farm), sharing data across labs (RoboNet), and in-person robotics competitions (DARPA Challenges, Amazon Picking Challenge). Any of these efforts alone cannot tackle all of the above, but they do illuminate a path forward: remotely-accessible robots on which everyone can run experiments, collect data, and benchmark their algorithms’ performance, the need for which was highlighted in the US Robotics Roadmap.


Robots in the cloud can:

1) Remove barriers to access by enabling remote experimentation on public hardware.

2) Allow for shared benchmarks through common environments and tasks.

3) Create large-scale robotic datasets through open-sourcing of recorded experiments.


Our workshop will examine the successes and challenges of robots in the cloud after 5 years of pioneering efforts in remote competitions and testbeds (Duckietown’s AI Driving Olympics, Georgia Tech's Robotarium, the Real Robot Challenge). We will bring together past and future leaders and participants of such efforts to discuss visions and open questions for the next 5 years. Workshop participants will include researchers and practitioners from academia, industry, and government across robotics, machine learning, and multi-institution research initiatives.

Schedule

Day: Friday, July 1st

  • [08:55 - 09:00] Opening Remarks

  • [09:00 - 10:30] Session 1: Benchmarking

    • [9:00 - 9:30] Lerrel Pinto | New York University

    • [09:30 - 10:00] Berk Calli | Worcester Polytechnic Institute

    • [10:00 - 10:30] Coline Devin | DeepMind

  • [10:30 - 11:00] Coffee Break

  • [11:00 - 12:30] Session 2: Pioneering Efforts

    • [11:00 - 11:30] Stefan Bauer | Max Planck Institute

    • [11:30 - 12:00] Liam Paull | Université de Montréal

    • [12:00 - 12:30] Sean Wilson | Georgia Tech

  • [12:30 - 14:00] Lunch

  • [14:00 - 15:30] Visions for the Future

    • [14:00 - 14:30] Carlotta Berry | Rose-Hulman

    • [14:30 - 15:00] Anca Dragan | UC Berkeley

  • [15:00 - 15:30] Coffee Break

  • [15:30 - 16:15] Speaker panel: open questions for the next 5 years

  • [16:15 - 16:45] Round tables: open questions

  • [16:45 - 17:00] Closing remarks (organizers)

Speakers


Organizers