Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts (ECCV 2026)

Official implementation of "Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts" (ECCV 2026).

Taewook Kang* | Taeheon Kim* | Donghyun Shin | Jonghyun Choi†

_{* Equal contribution    † Corresponding author}

[Paper] [Project Page]

🧭 Overview

TL;DR. Domain ARiThmetic (DART) adapts multi-task vision-language-action (VLA) models to environmental shifts using only one demonstration of one task through subspace-aligned weight arithmetic.

This repository provides:

• one-shot fine-tuning scripts for π₀.₅ on LIBERO visual-shift data,
• the DART implementation in domain_arithmetic/,
• LIBERO-view evaluation scripts.

📢 Updates

• Release the paper on arXiv.
• Launch the project page.
• Release code.
• Upload one-shot fine-tuning datasets and checkpoints.

📦 Installation

Clone the repository with submodules:

git clone --recurse-submodules git@github.com:snumprlab/dart.git
cd dart

If you already cloned the repository, initialize the submodules manually:

git submodule update --init --recursive

We use uv to manage Python dependencies. After installing uv, set up the environment with:

GIT_LFS_SKIP_SMUDGE=1 uv sync
GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .

GIT_LFS_SKIP_SMUDGE=1 is required because LeRobot is pulled as a dependency.

💾 Fine-tune VLA Models with One-Shot Data

Base VLA Model

We use the pi05_libero checkpoint from openpi as the source-trained multi-task VLA model. The checkpoint is trained on the LIBERO benchmark:

gs://openpi-assets/checkpoints/pi05_libero

By default, OpenPI downloads this checkpoint from gs://openpi-assets and caches it under ~/.cache/openpi. To change the cache location, set:

export OPENPI_DATA_HOME=/path/to/openpi_cache

Prepare the Dataset

One-shot fine-tuning expects a LeRobot-format LIBERO dataset. Set DATA_REPO_ID to either a local dataset path or a Hugging Face dataset repo id.

Supported third-person camera views are:

original (source), small, medium, large

For visual perturbations, use a separate dataset for each perturbation type.

Run One-Shot Fine-Tuning

Use exec/finetune.sh to fine-tune π₀.₅ on a selected camera view:

DATA_REPO_ID=/path/to/dataset \
CHECKPOINT_BASE_DIR=/path/to/checkpoints \
CAMERA_VIEW=medium \
CUDA_VISIBLE_DEVICES={GPU IDs} \
bash exec/finetune.sh

Common options:

Variable	Default	Description
CONFIG	pi05_libero_oneshotft	Training config
CAMERA_VIEW	original	Viewpoint shifts
BATCH_SIZE	64	Training batch size
NUM_TRAIN_STEPS	1000	Number of fine-tuning steps
SAVE_INTERVAL	1000	Checkpoint save interval

🎯 DART: One-Shot VLA Adaptation Through Domain Arithmetic

⚠️ Warning: The current DART implementation can require more than 100 GB of RAM. This may be reduced with further optimization.

DART requires three policies:

• base_policy: the original source-trained VLA model,
• policy_src: a one-shot policy fine-tuned in the source environment,
• policy_tgt: a one-shot policy fine-tuned in the target shifted environment.

The adapted policy is computed by applying the target-domain update after subtracting the source-specific component in an aligned subspace.

Run the example script:

bash exec/run_dart.sh

Or run DART directly:

uv run -m domain_arithmetic.dart \
  --cfg.base_policy.dir /path/to/base_policy \
  --cfg.base_policy.config pi05_libero \
  --cfg.policy_src.dir /path/to/source_oneshot_policy \
  --cfg.policy_src.config pi05_libero_oneshotft \
  --cfg.policy_tgt.dir /path/to/target_oneshot_policy \
  --cfg.policy_tgt.config pi05_libero_oneshotft \
  --cfg.scaling_coef 0.8 \
  --cfg.output_dir /path/to/output/DART_0.8 \
  --cfg.overwrite

The merged checkpoint is saved to --cfg.output_dir and can be evaluated with the same pi05_libero policy config.

🚀 Evaluation

Evaluation uses Docker to serve the policy and run LIBERO-view, a modified version of LIBERO for visual shifts.

Evaluate the default pi05_libero checkpoint:

GPU_ID=0 \
TASK_SUITE_NAME=libero_spatial \
CAMERA_VIEW=medium \
bash exec/eval.sh

Evaluate a DART checkpoint:

GPU_ID=0 \
POLICY_CONFIG=pi05_libero \
CHECKPOINT_DIR=/path/to/output/DART_0.8 \
TASK_SUITE_NAME=libero_spatial \
CAMERA_VIEW=medium \
bash exec/eval.sh

Add PERTURB argument to evaluate with visual perturbations:

GPU_ID=0 \
POLICY_CONFIG=pi05_libero \
CHECKPOINT_DIR=/path/to/output/DART_0.8 \
TASK_SUITE_NAME=libero_spatial \
CAMERA_VIEW=medium \
PERTURB=light_noise \
bash exec/eval.sh

Useful evaluation variables:

Variable	Default	Description
GPU_ID	0	Host GPU id used for evaluation
POLICY_CONFIG	pi05_libero	OpenPI policy config
CHECKPOINT_DIR	empty	Custom checkpoint path; empty uses the default LIBERO checkpoint
TASK_SUITE_NAME	libero_spatial	LIBERO task suite
CAMERA_VIEW	original	Evaluation camera view
PERTURB	empty	Optional visual perturbation: light or light_noise

📖 Citation

@inproceedings{kang2026dart,
  title     = {Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts},
  author    = {Kang, Taewook and Kim, Taeheon and Shin, Donghyun and Choi, Jonghyun},
  booktitle = {ECCV},
  year      = {2026},
}

🙏 Acknowledgements

This codebase is built on top of openpi and LIBERO.