Abstract
We present Interactive Gibson, the first comprehensive benchmark for trainingand evaluating Interactive Navigation: robot navigation strategies wherephysical interaction with objects is allowed and even encouraged to accomplisha task. For example, the robot can move objects if needed in order to clear apath leading to the goal location. Our benchmark comprises two novel elements:1) a new experimental setup, the Interactive Gibson Environment, whichsimulates high fidelity visuals of indoor scenes, and high fidelity physicaldynamics of the robot and common objects found in these scenes; 2) a set ofInteractive Navigation metrics which allows one to study the interplay betweennavigation and physical interaction. We present and evaluate multiplelearning-based baselines in Interactive Gibson, and provide insights intoregimes of navigation with different trade-offs between navigation pathefficiency and disturbance of surrounding objects. We make our benchmarkpublicly available(https://sites.google.com/view/interactivegibsonenv) andencourage researchers from all disciplines in robotics (e.g. planning,learning, control) to propose, evaluate, and compare their InteractiveNavigation solutions in Interactive Gibson.