Recoleta
GitHub
Source note

Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts

Authors
Taewook Kang, Taeheon Kim, Donghyun Shin, Jonghyun Choi
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionOne Shot AdaptationRobot Foundation ModelsWeight ArithmeticDomain Adaptation
Open arXivMarkdown source

Summary

DART adapts a Vision-Language-Action policy to a changed environment using one target-domain demonstration for one task. It extracts a domain-shift vector from fine-tuned weights and adds it to the base policy so other learned tasks can run in the target domain.

Problem

  • VLA policies can fail when camera pose, lighting, sensor calibration, or robot embodiment changes, even when the task itself stays the same.
  • Existing adaptation often needs demonstrations for each task in the target domain, which is expensive for real robot deployment.
  • One-shot fine-tuning on a single target task tends to improve that task while failing on held-out tasks, so it does not preserve the base policy’s multi-task behavior.

Approach

  • Fine-tune the base VLA model on one source-domain demonstration and one target-domain demonstration of the same task, producing two weight update vectors.
  • Subtract the source update from the target update to cancel the shared task direction and isolate a domain vector for the environmental shift.
  • Run SVD on layer updates and keep source singular directions that align with the target update subspace before subtraction, reducing source-domain artifacts.
  • Scale each layer’s domain vector by its source-target subspace alignment score, then add the scaled vector to the base model with a coefficient α.
  • The method changes weights only, so it does not require a new policy architecture.

Results

  • The excerpt gives no task-success percentages, standard errors, or baseline margins, so the main performance claim cannot be checked numerically from the provided text.
  • It claims one-shot adaptation with 1 target-domain demonstration for 1 task, plus a matching source-domain demonstration, then evaluation across all tasks in the source task set.
  • On LIBERO with π0.5 under a Medium camera-viewpoint shift, one-shot fine-tuning keeps higher performance on the adaptation task than on held-out tasks; no success rates are provided.
  • The weight analysis uses 16 update vectors from 4 tasks × 4 domains; the composed task-plus-domain estimate has the highest subspace alignment with target updates, but the excerpt gives no alignment values.
  • The paper reports experiments with π0.5 and π0-FAST in simulation and on real robots across visual and embodiment shifts, and claims DART beats existing one-shot VLA adaptation baselines without numeric margins in the excerpt.

Link