Technical Summary: Domain Arithmetic (DART) for One-Shot VLA Adaptation

Problem Statement

Vision-Language-Action (VLA) models, trained on large-scale corpora, exhibit strong multi-task capabilities within their source training environments. However, they suffer significant performance degradation when deployed in novel target environments characterized by environmental shifts. These shifts include changes in camera poses (viewpoint shifts), sensor calibrations, or embodiment modifications (e.g., switching from a Panda to a UR5e robot).

Existing adaptation strategies face a critical bottleneck:

1. Full-data Fine-tuning: Requires extensive expert demonstrations for every task in the target domain, which is prohibitively expensive to collect in real-world settings (e.g., household environments).
2. One-shot Fine-tuning: While data-efficient, standard fine-tuning on a single demonstration often fails to generalize to unseen tasks in the target domain. Instead of learning the environmental shift, the model primarily overfits to the specific task demonstrated, losing its multi-task capabilities.

Methodology: Domain ARiThmetic (DART)

The authors propose DART, an analogy-based framework that adapts VLA models to shifted environments using only a single demonstration of a single task. The method is grounded in the observation that one-shot fine-tuned weights can be decomposed into additive, disentangled components: task-specific directions and domain-specific directions.

Core Insight

Through subspace alignment analysis, the authors find that update-vectors ($\Delta = \theta_{fine-tuned} - \theta_{base}$) from one-shot fine-tuning are dominated by task-specific directions. However, a consistent, smaller domain-specific component exists that encodes the environmental shift. The authors hypothesize that these components are linearly additive and orthogonal, similar to the "queen = king + woman - man" analogy in language models.

Algorithmic Steps

DART extracts a reusable domain vector ($\tilde{\delta}_{tgt}$) to adapt the base policy ($\theta_0$) without retraining on large datasets.

1. Update-Vector Computation:

   • Compute the target-domain update-vector ($\Delta_{m,tgt}$) by fine-tuning the base model on a single target-domain demonstration ($D_{m,tgt}$) for task $T_m$.
   • Compute the source-domain update-vector ($\Delta_{m,src}$) by fine-tuning the same base model on a single source-domain demonstration ($D_{m,src}$) for the same task $T_m$.

2. Domain Vector Extraction (Analogy Operation):

   • Perform a subtraction to isolate the domain direction: $\delta_{tgt} = \Delta_{m,tgt} - \Delta_{m,src}$.
   • This operation cancels out the shared task-specific directions, leaving the domain-specific shift.

3. Subspace Filtering & Scaling (Noise Reduction):

   • Direct subtraction can retain source-domain artifacts or amplify fine-tuning noise. To address this, DART employs two refinement steps:
     • Subspace Filtering: Using Singular Value Decomposition (SVD), the method identifies and filters out basis vectors in the source update-vector that do not align with the target subspace. This prevents source-specific noise from corrupting the extracted domain vector.
     • Subspace Scaling: The resulting domain vector is scaled by the subspace alignment score ($\gamma$). If the source and target updates are poorly aligned (indicating high noise or irrelevance), the vector is down-weighted.

4. Model Adaptation:

   • The refined domain vector is added back to the base model: $\theta^* = \theta_0 + \alpha \cdot \tilde{\delta}_{tgt}$, where $\alpha$ is a scalar coefficient controlling adaptation strength.

Key Contributions

1. Empirical Evidence of Additivity: The paper provides evidence that one-shot fine-tuned weights in VLA models decompose into shareable, additive task- and domain-specific directions.
2. DART Framework: A novel analogy-based method that extracts a reusable domain vector from a single-shot fine-tuned weight update by removing task-specific directions via weight arithmetic.
3. Subspace Refinement: The introduction of subspace filtering and scaling to mitigate noise and artifacts during the extraction of the domain vector, ensuring robust adaptation.
4. Architecture Agnosticism: The method is demonstrated to work across different VLA architectures (flow-matching $\pi0.5$ and autoregressive $\pi0-FAST$) and diverse shift types (visual and embodiment).

Experimental Results

The authors evaluated DART on the LIBERO benchmark (simulation) and a real-world UR10e robot setup.

• Simulation (Novel Viewpoints): DART significantly outperformed baselines (Zero-shot, One-shot Fine-tuning, FLA, and RETAIN) across Small, Medium, and Large viewpoint shifts. For example, on $\pi0.5$ with Medium viewpoint shifts, DART achieved an 80.8% success rate compared to 33.3% for One-shot Fine-tuning and 76.4% for FLA.
• Combined Perturbations: DART maintained superior performance under combined visual shifts (Viewpoint + Noise + Light), demonstrating robustness to complex environmental changes.
• Cross-Embodiment: In transferring policies from a Panda robot to a UR5e robot, DART improved the success rate from 37.2% (Zero-shot) to 45.4%, and the progress rate to 73.8%, outperforming both One-shot Fine-tuning and Zero-shot baselines.
• Real-World: On a UR10e robot with a novel viewpoint, DART achieved an average success rate of 81.7% across five tasks (including fine-grained manipulation) using only a single demonstration for adaptation, significantly outperforming all baselines.
• Vector Merging: The authors demonstrated that domain vectors from different target domains can be merged into a single vector, allowing a single model to adapt to multiple environments simultaneously.

Significance and Claims

The paper claims that DART offers a data-efficient solution to the deployment bottleneck of VLA models. By leveraging the structural properties of weight updates, DART enables one-shot adaptation that preserves the multi-task capabilities of the base model while transferring it to new environments.

The authors emphasize that unlike prior approaches requiring extensive data or architectural modifications, DART is:

• Architecture-agnostic: Applicable to different VLA backbones.
• Hyperparameter-robust: Stable across a wide range of scaling coefficients ($\alpha$).
• Composable: Domain vectors can be merged to handle multiple shifts with a single parameter set.

The work suggests that the "domain knowledge" required for adaptation is implicitly learnable from a single demonstration and can be isolated mathematically, offering a practical path for deploying generalist robots in diverse, unstructured real-world settings.