InteriorSim: A Photorealistic Simulator for Embodied AI

Mike Roberts, Quentin Leboutet, Rachith Prakash, Renhan Wang, Hailin Zhang, Rui Tang, Marti Ferragut, Stefan Leutenegger, Stephan R. Richter, Vladlen Koltun, Matthias Müller, German Ros

Keywords: simulation environment, embodied AI

TL;DR: InteriorSim is a photorealistic simulator for embodied AI in the home.

Abstract: Interactive simulators are becoming powerful tools for training embodied agents, but existing simulators suffer from limited content diversity, physical interactivity, and visual fidelity. We address these limitations by introducing InteriorSim, a photorealistic simulator for embodied AI in the home. To create our simulator, we worked closely with a team of professional artists for over a year to construct 300 unique virtual indoor environments with 2,566 unique rooms and 17,234 unique objects that can be manipulated individually. Each of our environments features detailed geometry, photorealistic materials, and a unique floor plan and object layout designed by a professional artist, i.e., we do not rely on remixing existing layouts to create additional content. Our environments are implemented as Unreal Engine assets, and we provide an OpenAI Gym interface for interacting with the environments via Python. We demonstrate the utility of our simulator by using it in a zero-shot sim-to-real transfer scenario, i.e., we train a point-goal navigation policy entirely in simulation that can successfully navigate through cluttered real-world environments when deployed on a real robot. We also demonstrate that our simulator is quantitatively more photorealistic than existing simulators measured by human comparisons and standard metrics for evaluating generative models. Finally, we demonstrate that our simulator achieves better sim-to-real performance than existing simulators on a real-world semantic segmentation task. All of our assets and code will be made available online.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Infrastructure (eg, datasets, competitions, implementations, libraries)