Ken Goldberg, University of California, Berkeley

Ryan Hoque, University of California, Berkeley

Abstract

The very first robots, in Karel Čapek’s play R.U.R., acted collectively to rebel against unfair working conditions.  The first real robots, developed during WWII to handle radioactive materials, moved their mechanical arms under the close supervision of human ‘tele-operators’ who used levers behind shielded walls.

Since then, roboticists have assumed that robots must be self-contained and carry their own power supply, memory, and computing circuitry. However, over the past decade robots have started to collaborate again, with each other and with humans using advances in networking and cloud computing. Collaborative robotics has become a fast-growing sector of the market. All major robot companies FANUC, KUKA/Midea, ABB, and Omron Adept have introduced collaborative robots, as have new robot companies Universal Robots, Fetch, Franka Emika and Kinova.

At Amazon, Google, and other leading companies, fleets of robots contact remote human teleoperators when they are at risk or unable to make progress.  Fleet Learning is a new approach to human-robot collaboration that treats robots as novice learners and humans as their expert supervisors, where each human can supervise multiple robots. Input from the remote pool of humans can also be used to improve the robot fleet control policy over time. A central question is how to effectively allocate limited human attention, when multiple robots interactively query and learn from multiple human supervisors.

We’ll summarize very new results with an IFL metric and algorithms evaluated on an open-source benchmark suite of environments built on NVIDIA Isaac Gym with a fleet of 100 robots and physical experiments with 4 ABB YuMi robot arms and 2 remote humans. Experiments suggest that the allocation of humans to robots can significantly affect the performance of the fleet, and that the new algorithm can achieves up to 8.8x higher return on human effort than baselines.