Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts

Illustration of our proposed Domain Arithmetic (DART).

TL;DR

Domain ARiThmetic (DART) adapts multi-task VLAs to environmental shifts (e.g., camera-pose changes, embodiment changes) using a single demo of a single task through subspace-aligned weight arithmetic.

Motivation

VLA Models are fragile under Environmental Shifts

Vision-Language-Action (VLA) models generate robot actions from camera images and language instructions. While they can solve many tasks in their training environment (source domain), environmental shifts (e.g., camera-pose shifts or embodiment changes) can make them fail in a new environment (target domain).

DART teaser figure — Environmental shifts can make a source-trained VLA model fail. Full-data fine-tuning is costly, while one-shot fine-tuning often fails to generalize beyond the demonstrated task. This motivates one-shot adaptation to new environments.

Adapting VLA Models from Source to Target Domains

Full-data Fine-tuning is Costly

Requires target-domain demonstrations for every task.

One-shot Fine-tuning is Narrow

Adapts to demo tasks, but fails to generalize across diverse tasks.

DART transfers Domain Knowledge

Extracts domain knowledge from one demo and applies it across tasks.

Fine-tuned Weights Analysis

Task and Domain Components in One-Shot Updates

One-shot updates mainly capture task knowledge, but also contain reusable domain components. These task and domain components compose additively within one-shot fine-tuning updates.

Additive alignment analysis — One-shot fine-tuning captures both task and domain components. Adding them closely matches the actual model update.

Domain heatmap analysis — Similar environmental shifts produce aligned weight updates, while combined shifts partially reuse update components from each shift.

Finding #1: Task-Domain Decomposition

Task and domain components can be combined to approximate the actual one-shot update, suggesting a decomposable structure in weight space.

Finding #2: Structured Domain Components

Environmental shifts produce structured weight updates: similar shifts lead to similar updates, while combined shifts reuse patterns from individual shifts.

Method

Domain Arithmetic (DART) for One-Shot Adaptation

DART adapts a multi-task VLA model to a new environment using just one demonstration of a single task, instead of collecting demonstrations for every task.

DART method overview — DART adapts the source-domain multi-task VLA model to the target domain using one-shot data by using a target-domain component (vector): it subtracts the source-domain update from the target-domain update to remove shared task components.

1 Fine-tune in Source & Target Domains

Fine-tune the source-domain multi-task VLA model θ₀ on one same-task demo from each domain.

2 Compute Update Vectors

Measure how the multi-task VLA model θ₀ changes in the source and target domains (update vectors).

3 Remove Task-Specific Change

Filter the source vector, then subtract it from the target vector to extract the domain vector.

4 Adapt VLA Model to Target Domain

Add the extracted domain vector to the VLA model θ₀ to obtain the target-domain multi-task policy θ^*.

Experiments

Experimental Setup

Quantitative Results

DART consistently outperforms baselines under the one-shot setting across diverse simulation and real-world shifts.

Novel Visual Domains

On LIBERO, DART achieves the best average success under novel viewpoints with both π_0.5 and π₀-FAST, and also outperforms baselines under visual perturbations with π_0.5.

Cross-Embodiment Transfer

On MimicGen, DART adapts a Panda-trained policy to UR5e using a single Stack demonstration, improving average success to 69.4% across Stack and Stack Three.

Real-World Experiments

Using a single Stack Cube demonstration, DART adapts to the target viewpoint and achieves 81.7% average success rate across five real-world UR10e tasks.

DART experimental results — Performance (success rate) on LIBERO tasks across diverse visual shifts, including novel viewpoints and combined camera noise / illumination changes with π_0.5 and π₀-FAST. See the paper for detailed numbers.

Stack task progress and success rate table — Performance on MimicGen tasks under Panda-to-UR5e cross-embodiment transfer with π_0.5, using only a single Stack demonstration.

Novel viewpoint success rate table — Performance on real-world UR10e manipulation tasks under novel viewpoints with π_0.5, using only a single Stack Cube demonstration.

BibTeX

@inproceedings{kang2026dart,
  title     = {Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts},
  author    = {Kang, Taewook and Kim, Taeheon and Shin, Donghyun and Choi, Jonghyun},
  booktitle = {ECCV},
  year      = {2026},
}