Multimodal / Robotics / Training
Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts
** Taewook Kang, Taeheon Kim, Donghyun Shin, Jonghyun Choi (Seoul National University)
Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts
Authors: Taewook Kang, Taeheon Kim, Donghyun Shin, Jonghyun Choi (Seoul National University)
arXiv ID: 2607.00666
Problem: Vision-Language-Action (VLA) models fail under environmental shifts (camera pose, embodiment changes) and existing adaptation methods require costly multi-demonstration data per task.
Key Methodology:
- DART (Domain ARiThmetic): An analogy-based weight arithmetic approach that extracts a "domain vector" by subtracting a source-domain update-vector from a target-domain update-vector (both fine-tuned from a single demonstration of one task), then adds this vector to the base model - transferring multi-task capabilities to the target domain without per-task retraining.
- Subspace Filtering: Filters misaligned singular components between source and target update-vectors to suppress noise and source-domain artifacts, using SVD-based overlap energy and a dynamic cutoff derived from subspace alignment scores.
- Subspace Scaling: Down-weights domain vectors by the alignment score when source-target subspaces are fundamentally misaligned, modulating noisy or irrelevant contributions.
Key Results:
- On LIBERO with π₀.5, DART achieves 79.1% avg. success rate across Small/Medium/Large viewpoint shifts vs. 74.3% (FLA), 69.6% (RETAIN), 31.5% (One-shot FT), and 54.5% (Zero-shot) - a +24.6 pp gain over zero-shot.
- Under combined visual perturbations (View+Noise+Light): 75.0% vs. 71.5% (FLA) and 60.5% (Zero-shot).
- Cross-embodiment (Panda → UR5e): 69.4% avg. success rate vs. 62.0% (Zero-shot) and 56.4% (One-shot FT).
- Real-world UR10e robot with novel viewpoint: 81.7% avg. success across 5 tasks vs. 55.0% (FLA) and 43.3% (Zero-shot), using only a single Stack Cube demo for adaptation.
Applied Context: DART lets you adapt a VLA policy to a new camera setup or robot platform with literally one demonstration of one task - no per-task data collection, no architecture changes. This makes deploying base models into new physical environments (different labs, different camera rigs, different robot arms) practical and data-efficient.