Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts¶

🕒 Published (v1): 2026-07-01 09:13 UTC · Source: Arxiv · Venue: ECCV 2026 · link

Why this paper was selected

ECCV 2026; domain arithmetic for VLA one-shot adaptation, directly relevant to robot VLM deployment

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

DART (Domain ARiThmetic) addresses one-shot adaptation of Vision-Language-Action (VLA) models to new environments by performing weight-space analogy operations. It extracts a reusable domain vector by subtracting a source-domain update from a target-domain update, then applies SVD-based subspace filtering and scaling to suppress noise before adding the domain vector back to the base policy. The result is a parameter-efficient, architecture-agnostic method that transfers multi-task capabilities to novel environments from a single demonstration.

Problem¶

VLA models trained in a source environment degrade substantially under environmental shifts (camera pose, embodiment changes, lighting, noise). Existing adaptation methods require many task-wise demonstrations in the target domain, creating prohibitive data collection costs. One-shot fine-tuning on a single target demonstration fails to generalize to held-out tasks because the resulting weight update is dominated by task-specific rather than domain-specific directions.

Method¶

DART is built on the empirical finding that a one-shot update-vector $\Delta_{m,\text{tgt}} = \theta_{m,\text{tgt}} - \theta_0$ decomposes additively into orthogonal task-specific and domain-specific directions. To extract the domain direction:

Update-vector computation. Fine-tune the base policy $\theta_0$ on one demonstration each from the source domain ($D_{m,\text{src}}$) and the target domain ($D_{m,\text{tgt}}$), yielding $\Delta_{m,\text{src}}$ and $\Delta_{m,\text{tgt}}$.
Subspace filtering. For each layer $l$, compute SVDs of both update-vectors. Form the interaction matrix $C^{(l)} = U_{\text{tgt}}^{(l)\top} U_{\text{src}}^{(l)}$ and rank source basis vectors by overlap energy $e_j^{(l)} = \|C^{(l)}_{:,j}\|_2^2$. Retain only the top-$r_l$ source basis vectors whose cumulative energy meets the subspace alignment threshold $\gamma^{(l)}$, producing a filtered source update $\tilde{\Delta}_{m,\text{src}}^{(l)}$.
Subspace scaling and domain vector. Compute the refined domain vector per layer: $$\tilde{\delta}_{\text{tgt}}^{(l)} = \gamma^{(l)} \cdot \left(\Delta_{m,\text{tgt}}^{(l)} - \tilde{\Delta}_{m,\text{src}}^{(l)}\right)$$ The alignment score $\gamma^{(l)}$ down-weights layers where source and target subspaces are fundamentally misaligned (noise-dominant).
Adaptation. Inject the domain vector into the base policy: $\theta^* = \theta_0 + \alpha \cdot \tilde{\delta}_{\text{tgt}}$, where $\alpha$ is a scalar strength coefficient. The multi-task capabilities of $\theta_0$ are preserved; only environmental adaptation is added.

Key Contributions¶

Empirical demonstration that one-shot VLA update-vectors decompose into additive, orthogonal task- and domain-specific directions, validated via subspace alignment scores across 16 update-vectors (4 tasks × 4 domains).
DART framework: an analogy-based weight arithmetic method that extracts a reusable domain vector from one-shot fine-tuned weights without architectural changes or additional data beyond a single paired demonstration.
Subspace filtering and scaling: SVD-based techniques that remove cross-domain noise before subtraction, improving domain vector purity.
Architecture-agnostic: evaluated on both flow-matching ($\pi_{0.5}$) and autoregressive ($\pi_0$-FAST) VLA backbones.
Enables domain merging across multiple target environments simultaneously.

Results¶

On LIBERO (visual viewpoint shifts — Small/Medium/Large — and perturbation combinations including noise and lighting), DART outperforms all one-shot baselines (Zero-shot, One-shot FT, FLA, RETAIN) across total, adaptation, and held-out tasks.
On MimicGen cross-embodiment (Panda → UR5e), DART again exceeds all baselines under one-shot scenarios.
Real-world experiments with two third-person viewpoints confirm the approach transfers beyond simulation.
Evaluated with π0.5 as primary model and π0-FAST to confirm architectural generality; both show consistent gains.
Specific headline numbers are not fully extracted in the provided text excerpt (tables are referenced but not reproduced), but the paper states DART "outperforms existing VLA adaptation methods in one-shot scenarios across diverse visual and embodiment shifts."

Limitations¶

Requires a source-domain demonstration of the same adaptation task, which must be available from the original training data; tasks without source coverage cannot be adapted.
The scalar coefficient $\alpha$ controlling adaptation strength requires selection, although the paper claims hyperparameter robustness.
Evaluation is limited to manipulation tasks (LIBERO, MimicGen); generalization to navigation or other VLA settings is unvalidated.
One-shot data collection in the target domain is still required (one demonstration per environment, not zero-shot).
Domain vector extraction via subtraction inherently depends on the assumption that source and target fine-tuned models share task directions; violations (e.g., severe task-domain entanglement) could degrade the extracted vector.

Relevance to Vision-Language Models¶

DART directly extends weight arithmetic techniques—previously explored in vision-language models for task composition—to the VLA setting, demonstrating that the same additive, disentangled structure observed in VLM fine-tuning transfers to embodied action policies. The subspace alignment analysis methodology is broadly applicable to understanding how VLMs encode different semantic concepts in weight space. For researchers tracking VLMs, this work provides evidence that the low-rank, compositional structure of fine-tuned weight updates is a general property of large multimodal models, opening the door to analogous adaptation techniques for VLMs operating under distribution shift (e.g., domain transfer in visual question answering or cross-camera visual grounding).