Introduction

The Human Input Parsing Platform for Openai Gym (HIPPO Gym) provides a platform for Human-AI interaction research with geographically distributed research participants. Designed for rapid deployment and iteration, it is hoped that HIPPO Gym can let researchers focus on collecting and utilizing Human Input data without the time-consuming process of having to write custom code for each research project.

Written by: Nick Nissen (opens new window), Yuan Wang, Nadeen Mohamed (opens new window), and Payas Singh (opens new window)

Supervised by: Matt Taylor (opens new window) and Neda Navi

For the Intelligent Robot Learning Laboratory (IRLL) (opens new window) at the University of Alberta (UofA) (opens new window)

Supported by the Alberta Machine Intelligence Institute (AMII) (opens new window)

Overview

For the convenience of RL Researchers, HIPPO Gym is written in Python, and utilizes simple JSON for messaging, both tools that should be familiar to most RL Researchers.

The overall system consists of mostly on-demand infrastructure setup in AWS, a relatively simple front-end that can service multiple research projects within an organization, and a Docker-deployed server that connects the front-end to a research-project-specific agent/environment combo.

The server is designed to be ephemeral, it is created only when a research participant starts the first-step of a research project (ie lands on the consent page), and the server ceases existence at the end of that participant's task. This system provides significant cost advantages as well as providing one-server-per-participant guarantees that the task of one participant cannot impact those of another participant.

In order to facilitate Researchers, once an organization's AWS setup has been completed, individual researchers can deploy new or updated research projects by simply adjusting a config.yml file for their specific use case and then running a deployment script. This process takes less than 5 minutes on average, not counting the time it takes to develop an agent of course.