Hardware Realization of Long-Horizon Humanoid Interaction

Establish a hardware-capable humanoid robotic system that can fluidly execute long-horizon sequences of heterogeneous interaction tasks—such as pushing a chair aside, picking up a box, carrying it across a room, and sitting down—mirroring skilled human performance across varied contact patterns, object geometries, and body configurations.

Background

The paper motivates the problem with a household tidying scenario that involves sequential, contact-rich tasks requiring distinct contact patterns and body configurations. Although recent work has advanced locomotion and dynamic motion control for humanoids, translating these skills into persistent, contact-rich interaction with varied objects remains challenging.

The authors frame the difficulty as primarily representational: reference-based approaches tie policies to demonstration geometry and struggle to adapt, whereas reference-free methods often rely on task-specific designs and fail to generalize or compose across skills. Achieving the human-like capability of executing long sequences of heterogeneous tasks on real hardware is identified as an outstanding challenge.