Mobile Object Rearrangement with Learned Localization Uncertainty

Wenyu Han, Rongjing Xie, Ashiq Rahman Anwar Batcha, Chen Feng

Primary Area: applications to robotics, autonomy, planning

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.

Keywords: Mobile object rearrangement, Uncertainty estimation, Reinforcement learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: DRL-compatible localization uncertainty estimation improves mobile rearrangement

Abstract:

Mobile object rearrangement (MOR) is a pivotal embodied AI task for a mobile agent to move objects to their target locations.
While previous works rely on accurate pose information, we focus on scenarios where the agent needs to always localize both itself and the objects. This is challenging because accurate rearrangement depends on precise localization, yet localization in such a non-static environment is often disturbed by changes in the surroundings after rearrangement. To address this challenge, we first learn an effective representation for MOR only from sequential first-person view RGB images. It recurrently estimates agent and object poses, along with their associated uncertainties. With such uncertainty-aware localization as the input, we can then hierarchically train rearrangement policy networks for MOR. We develop and open source a simplified, yet challenging 3D MOR simulation environment to evaluate our method and relevant embodied AI baselines. Extensive comparisons reveal better performances of our method than baselines and the need for uncertainty estimation in our task.

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.

Submission Number: 6758