Achieving long-horizon non-prehensile rearrangement under partial observability

Develop algorithms that enable mobile robots to perform long-horizon non-prehensile rearrangement using only egocentric sensing under genuinely partial observability, characterized by narrow fields of view, frequent occlusions, and the absence of global state information.

Background

The paper studies multi-object non-prehensile rearrangement from egocentric vision, where objects frequently occlude each other and global state estimation is unreliable. Existing approaches often assume global state or external tracking and struggle when only partial, egocentric observations are available.

In discussing prior work, the authors explicitly identify a remaining open challenge: executing long-horizon non-prehensile rearrangement in settings with genuinely partial observability, where the robot must rely solely on limited egocentric cues without access to global maps or consistent localization. EgoPush is proposed as a step toward this goal by learning object-centric representations and constraining a privileged teacher to egocentric, visibility-limited observations.