Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts

2026-07-01

AI Summary

This work addresses the poor generalization of existing Vision-Language-Action (VLA) models under environmental shifts—such as changes in camera viewpoint or robot embodiment—and the reliance of conventional adaptation methods on extensive task demonstrations. To overcome these limitations, the authors propose a single-shot, cross-domain adaptation approach grounded in analogical reasoning, which enables rapid transfer using only one demonstration. The method leverages singular value decomposition of weight vectors to construct a subspace alignment mechanism that precisely extracts and injects domain-specific information while effectively suppressing noise. Experimental results demonstrate that this approach significantly outperforms current VLA adaptation techniques in both simulated and real-world settings across tasks involving visual and proprioceptive domain shifts, all within a single-shot adaptation paradigm.

Abstract

Vision-Language-Action (VLA) models often fail to perform the same learned tasks under environmental shifts, such as changes in camera pose and shifts to a different but similar robot (e.g., from Panda to UR5e). Adapting these models to the shifted environment (i.e., target domain) often requires training on multiple demonstrations for each task, which are costly to collect. To reduce the burden of data curation and training, we propose an analogy-based method that adapts VLA models under environmental shifts through weight vector arithmetic with domain-specific information addition, named Domain ARiThmetic (DART). Unlike prior approaches, DART requires collecting only a single demonstration, enabling efficient adaptation. To accurately isolate domain-specific information for addition, DART performs subspace alignment between singular components in weight vectors to filter out noisy components. In both simulated and real-world experiments, DART outperforms existing VLA adaptation methods in one-shot scenarios across diverse visual and embodiment shifts. Code is available at https://github.com/snumprlab/dart.

Problem

• Vision-Language-Action
• environmental shifts
• one-shot adaptation
• domain adaptation
• robot embodiment

Innovation

• Domain Arithmetic
• One-Shot Adaptation
• Vision-Language-Action
• Subspace Alignment
• Environmental Shift