Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts

This paper presents Domain ARiThmetic (DART), an analogy-based method for adapting vision-language-action (VLA) models to environmental shifts with only one target-domain demonstration. DART adds domain-specific information through weight-vector arithmetic and uses subspace alignment between singular components to filter noisy components before addition. The authors evaluate the method under visual shifts, such as camera-pose changes, and embodiment shifts, such as moving from a Panda robot to a UR5e. According to the paper abstract, DART outperforms existing VLA adaptation methods in one-shot settings across simulated and real-world experiments. Code is available from the authors’ GitHub repository.