Taewook Kang, Taeheon Kim, Donghyun Shin, Jonghyun Choi
4 min
Vision-Language-Action (VLA) models often struggle when deployed in environments that differ from their training data, such as changes in camera perspective or robot embodiment. Traditional fine-tuning requires extensive expert demonstrations for every task in the new environment, which is costly and impractical. This paper asks: can we adapt VLA models to new environments using only a single demonstration by leveraging the structural properties of weight updates?
The authors propose Domain ARiThmetic (DART), a framework that treats model adaptation as an analogy operation. By analyzing one-shot fine-tuned weights, the researchers discovered that parameter updates can be decomposed into additive task-specific and domain-specific directions. DART extracts the domain-specific component by subtracting a source-domain update-vector from a target-domain update-vector. To prevent noise and task-specific artifacts from corrupting this domain vector, the authors introduce subspace filtering—which aligns and retains only shared singular components—and subspace scaling to down-weight irrelevant noise.
DART consistently outperforms existing one-shot adaptation methods across diverse simulated and real-world benchmarks, including viewpoint shifts, camera noise, and cross-embodiment transfers (e.g., Panda to UR5e). The method is architecture-agnostic, showing success with both flow-matching and autoregressive VLA models. Furthermore, the domain vectors extracted by DART are composable; multiple domain vectors can be merged into a single, transferable vector, allowing for efficient adaptation across multiple target environments without storing separate models.
This work provides a highly data-efficient path for deploying robotic foundation models in the real world. By reducing the adaptation requirement to a single demonstration, DART significantly lowers the barrier for deploying VLA models in novel, dynamic settings like household environments, where collecting large-scale task-specific data is infeasible.
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.
Alex: So the real innovation is figuring out how to separate the "what"—the task—from the "where"—the environment—cleanly enough that the math actually works.
Sam: Precisely. And the practical implication is significant. Instead of collecting dozens or hundreds of examples in a new room and retraining from scratch, the robot can adapt from a single demonstration. The paper suggests this works because neural networks—the systems that power modern robots—aren't just black boxes. They have internal structure that we can reason about and, to some extent, do arithmetic on.
Alex: That's a meaningful shift in how we think about deploying robots in the real world. Rather than treating every new environment as a completely new problem, DART treats it as a small, calculable adjustment to an existing solution.
Sam: That's a fair summary. It's worth noting the researchers are careful about the scope of their claims—this approach works within the framework they've tested, and there will be environments or tasks where the separation between "task" and "domain" isn't as clean. But as a proof of concept, it opens up a useful direction: that adapting a robot to a new place might one day be as simple as showing it around the room once.
Alex: A single walk-through instead of weeks of retraining. Thanks for walking me through that, Sam, and thanks to everyone listening to ResearchPod.