Unified Interaction Representation for Reference-Free, Generalizable Multi-Skill Control

Develop a unified interaction representation for humanoid–object interaction that enables reference-free inference, generalizes across object geometries and scales, and supports long-horizon skill composition within a single policy.

Background

The paper argues that existing methods fall into two camps: reference-based methods that achieve fidelity but overfit to demonstrated geometries, and reference-free methods that require task-specific rewards and lack cross-task unification. This motivates the search for a single representation that can support autonomy and generalization without relying on motion references.

The authors propose distance fields as a potential solution, but preface their contribution by stating that identifying such a representation that simultaneously meets all three desiderata—reference-free inference, geometric generalization, and long-horizon skill composition within one policy—has been an open challenge.