CAREER: Crowdsourcing for Multirobot Coordination

Brown University, Providence RI

Investigators

Abstract

Teams of humans are exceptionally good at coordination. Teams of robots, however, are extremely clumsy at coordination, requiring extensive communication and computation. Reliance on this infrastructure poses a significant roadblock to bringing robot teams into real-world applications. This project is pursuing an integrated research, education, and outreach approach for developing novel, data-driven algorithms for multi-robot coordination, inspired by human coordination. As cognitive beings that make decisions based on broad context, memory, and sensing, human capabilities are challenging to transfer to robotics. To facilitate this transfer, the project is developing an online crowdsourcing application that tasks participants with creating a global structure, such as a shape. The application constrains participants to robot-like capabilities by limiting available information and actions. The application will provide a faithful representation of the capabilities of distributed teams of robots, and will be used to gain insights into human coordination that can then be transferred to a multi-robot system. The overarching goal of the proposed work is to develop novel methodologies for multi robot coordination firmly grounded in human collaboration, based on models learned from data collected via a crowdsourced online application. To this end, the research objectives are (1) to explicate the relationship between context (communication and sensing) and outcomes in distributed teams of humans working on tightly coupled tasks using data generated from an online multi-person interface; (2) to identify, using statistical methods, parameters for distributed teams of robots solving similar shared objective problems; (3) to infer, using deep learning architectures, diverse ensembles of coordination models for distributed teams of robots solving tightly coupled problems using the data collected from the crowdsourcing application; and (4) to validate these models by evaluating their success in solving tightly coupled problems using a combination of simulation, hardware, and mixed reality experiments.