pith. sign in

arxiv: 2607.00666 · v1 · pith:JBRLKESMnew · submitted 2026-07-01 · 💻 cs.RO · cs.CV· cs.LG

Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts

Taewook Kang , Taeheon Kim , Donghyun Shin , Jonghyun Choi This is my paper

Pith reviewed 2026-07-02 11:33 UTC · model grok-4.3

classification 💻 cs.RO cs.CVcs.LG
keywords vision-language-actionone-shot adaptationdomain adaptationweight vector arithmeticroboticsembodiment shiftvisual shiftsubspace alignment
0
0 comments X

The pith

VLA models adapt to new camera views or robot bodies using one demonstration by adding isolated domain information via weight arithmetic.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents DART as a way to adapt vision-language-action models when the setting changes, such as a shifted camera or a switch to a similar but different robot arm. Standard adaptation needs several demonstrations per task, but DART reduces this to a single demonstration by isolating domain-specific details from the weights and adding them through vector arithmetic. Subspace alignment on singular components filters out noise so the addition stays clean. This matters because demonstration collection is costly in time and resources, making one-shot methods more usable for robots operating in varied conditions. If the approach holds, VLA systems become practical to retarget without repeated full retraining.

Core claim

DART adapts pretrained VLA models to a target domain under environmental shifts by collecting one demonstration, extracting domain-specific information through subspace alignment between singular components in weight vectors, and incorporating that information via weight vector arithmetic.

What carries the argument

Domain ARiThmetic (DART), which isolates domain-specific information via subspace alignment of singular components in weight vectors before performing addition through vector arithmetic.

If this is right

  • Only one demonstration per task is needed for adaptation instead of multiple.
  • The method outperforms prior one-shot VLA adaptation techniques in both simulation and real-world tests.
  • It handles both visual shifts like camera pose changes and embodiment shifts like switching between similar robot arms.
  • No full retraining on target-domain data is required beyond the single demonstration.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The subspace alignment step could be tested on other weight-based adaptation problems outside VLA models.
  • If the arithmetic generalizes, similar one-shot methods might apply to transferring policies across different sensor setups.
  • Reducing data collection to one example could allow faster iteration when deploying VLA systems in new physical layouts.

Load-bearing premise

That subspace alignment between singular components in weight vectors from a single demonstration can accurately separate domain-specific information from task content and noise.

What would settle it

In controlled trials on new visual or embodiment shifts, the one-shot DART-adapted model performs no better than the unadapted baseline or falls short of models trained on multiple demonstrations from the target domain.

Figures

Figures reproduced from arXiv: 2607.00666 by Donghyun Shin, Jonghyun Choi, Taeheon Kim, Taewook Kang.

Figure 1
Figure 1. Figure 1: One-shot VLA adaptation under environmental shifts. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Properties of one-shot fine-tuning. (a) The model is fine-tuned on adapta￾tion tasks in target (Medium) camera viewpoint. Performance remains high in adaptation tasks but generalizes poorly to other held-out tasks. (b) Subspace alignment γ(·, ·) among update-vectors ∆m,tgt = θm,tgt − θ0 on m ∈ {1, 2, 3} and tgt ∈ {Source, Medium}. Vec￾tors align for the same task and domain, showing task- and domain-shared… view at source ↗
Figure 3
Figure 3. Figure 3: Additive task-domain directions in update-vectors. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the proposed VLA adaptation approach. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of experimental setups. We experiment on four setups: simulation setups with novel viewpoints (top-left) and combined visual perturbations (bottom-left) on LIBERO [37], a cross-embodiment transfer setup on MimicGen [42] (middle), and a real-world setup on two third-person camera viewpoints (right). 6 Experiments To evaluate our method, we conduct (i) simulation experiments under diverse visual shi… view at source ↗
Figure 6
Figure 6. Figure 6: Performance under hyperparameter choices on LIBERO across novel [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of feature shifts in￾duced by task and domain prototypes. B Details on Baseline Methods B.1 RETAIN [67] RETAIN is a parameter merging method for VLA models that enables learning a new task while mitigating forgetting of previously learned tasks. It interpolates [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example rollouts on MimicGen. Top: Example rollout of the Stack task with UR5e. Bottom: Example rollout of the Stack Three task with UR5e [PITH_FULL_IMAGE:figures/full_fig_p030_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Real-world setup and example rollouts. Left: [PITH_FULL_IMAGE:figures/full_fig_p031_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Impact of scaling coefficient α. 25 50 75 100 125 Training Time (Minutes) 72 75 78 81 Success Rate (%) FLA RETAIN DART [PITH_FULL_IMAGE:figures/full_fig_p038_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Impact of choice of layers to adapt. Vis is vision encoder, LLM is language model, and Action is action expert in the VLA model, π0.5. We report the average success rate (%) on LIBERO in three novel viewpoint shifts (Small, Medium, Large) (Left). We also measure the average absolute value of the domain vectors across the chosen layers (Right). that viewpoint shifts affect not only perception but also lang… view at source ↗
Figure 14
Figure 14. Figure 14: Per-layer statistics of DART on LIBERO across novel viewpoints in π0.5. We plot the mean and standard deviation across three novel viewpoints (Small, Medium, Large) and three different adaptation tasks Tm, m ∈ {1, 2, 3}. VIS is a vision encoder, LLM is a language model, ACTION is an action expert in the VLA model. alignment score and the overlap energy. As we can naturally expected from the trends observe… view at source ↗
read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Domain ARiThmetic (DART), a one-shot adaptation method for Vision-Language-Action (VLA) models under environmental shifts (e.g., camera pose changes or embodiment shifts like Panda to UR5e). It derives a domain vector via weight arithmetic from a single demonstration and applies subspace alignment using SVD on singular components to isolate domain-specific information before addition. Experiments in simulation and real-world settings claim DART outperforms prior VLA adaptation methods across visual and embodiment shifts. Code is released at https://github.com/snumprlab/dart.

Significance. If the central claims hold after addressing validation gaps, the work would meaningfully lower the data-collection cost for deploying VLA models in new environments, a practical bottleneck in robotics. The explicit release of code is a positive contribution to reproducibility.

major comments (2)
  1. [Method] Method section (DART derivation and subspace alignment): the claim that SVD-based alignment on singular components isolates domain factors from a single demonstration rests on an untested separability assumption. No analysis is provided of singular-spectrum stability across multiple single demonstrations of the same task, nor a control experiment showing that alignment removes trajectory/gripper/visual noise rather than useful signal. This directly undermines the one-shot isolation guarantee.
  2. [Experiments] Experiments section: the reported outperformance in one-shot scenarios is presented without ablations that isolate the contribution of the alignment step versus plain weight arithmetic, or without reporting variance across different choices of the single demonstration. This makes it impossible to confirm that gains arise from the proposed filtering rather than from other implementation details.
minor comments (2)
  1. [Method] Notation for the domain vector and the SVD truncation threshold should be defined explicitly with an equation number rather than described in prose only.
  2. [Experiments] Figure captions for the real-world robot setups should include the exact number of trials and success criteria to allow direct comparison with baselines.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their insightful comments on our manuscript. We address the major concerns point-by-point below, proposing revisions to enhance the clarity and rigor of our claims regarding DART's one-shot adaptation method.

read point-by-point responses

  1. Referee: [Method] Method section (DART derivation and subspace alignment): the claim that SVD-based alignment on singular components isolates domain factors from a single demonstration rests on an untested separability assumption. No analysis is provided of singular-spectrum stability across multiple single demonstrations of the same task, nor a control experiment showing that alignment removes trajectory/gripper/visual noise rather than useful signal. This directly undermines the one-shot isolation guarantee.

    Authors: The subspace alignment in DART is derived from the principle that domain shifts manifest as consistent directions in the weight space, separable via SVD from task-specific components. While we did not include explicit stability analysis in the original submission, the empirical success across diverse shifts supports the practical utility of the approach. We will revise the manuscript to include an analysis of singular spectrum stability using multiple single-demonstration examples and a control experiment to verify that the alignment primarily filters noise rather than signal. revision: yes

  2. Referee: [Experiments] Experiments section: the reported outperformance in one-shot scenarios is presented without ablations that isolate the contribution of the alignment step versus plain weight arithmetic, or without reporting variance across different choices of the single demonstration. This makes it impossible to confirm that gains arise from the proposed filtering rather than from other implementation details.

    Authors: We agree that ablations are necessary to attribute performance gains specifically to the subspace alignment. In the revised manuscript, we will add an ablation study comparing DART with and without the SVD-based alignment, as well as report standard deviations or variance metrics across different selections of the single demonstration used for domain vector computation. revision: yes

Circularity Check

0 steps flagged

No circularity: method is a direct procedural construction without self-referential reduction

full rationale

The paper presents DART as a procedural adaptation technique that extracts a domain vector from one demonstration via weight arithmetic, followed by SVD-based subspace alignment to filter components before addition. No equations, derivations, or claims in the provided text reduce a 'prediction' or result to the input by construction (e.g., no parameter fitted on a subset then renamed as a prediction of a related quantity). No self-citations are invoked as load-bearing for uniqueness or ansatz. The central claim rests on an empirical assumption about singular vector separability, which is an untested modeling choice rather than a definitional loop. This qualifies as a self-contained method proposal with no circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities explicitly mentioned or derivable from the abstract.

pith-pipeline@v0.9.1-grok · 5720 in / 919 out tokens · 25127 ms · 2026-07-02T11:33:21.649695+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

80 extracted references · 10 canonical work pages · 6 internal anchors

  1. [1]

    arXiv preprint arXiv:2509.14117 (2025)

    Abouzeid, A., Mansour, M., Sun, Z., Song, D.: Geoaware-vla: Implicit geometry aware vision-language-action model. arXiv preprint arXiv:2509.14117 (2025)

    work page arXiv 2025

  2. [2]

    PaliGemma: A versatile 3B VLM for transfer

    Beyer, L., Steiner, A., Pinto, A.S., Kolesnikov, A., Wang, X., Salz, D., Neumann, M., Alabdulmohsin, I., Tschannen, M., Bugliarello, E., et al.: Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    Bjorck, J., Castañeda, F., Cherniadev, N., Da, X., Ding, R., Fan, L., Fang, Y., Fox, D., Hu, F., Huang, S., et al.: Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    In: CoRL (2025)

    Black, K., Brown, N., Darpinian, J., Dhabalia, K., Driess, D., Esmail, A., Equi, M.R., Finn, C., Fusai, N., Galliker, M.Y., et al.:π0.5: A vision-language-action model with open-world generalization. In: CoRL (2025)

    2025

  5. [5]

    RSS (2025)

    Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., et al.:π0: A vision-language-action flow model for general robot control. RSS (2025)

    2025

  6. [6]

    In: CoRL (2024)

    Chen, L.Y., Xu, C., Dharmarajan, K., Irshad, M.Z., Cheng, R., Keutzer, K., Tomizuka, M., Vuong, Q., Goldberg, K.: Rovi-aug: Robot and viewpoint augmenta- tion for cross-embodiment robot learning. In: CoRL (2024)

    2024

  7. [7]

    NeurIPS (2019)

    Chen, X., Wang, S., Fu, B., Long, M., Wang, J.: Catastrophic forgetting meets negative transfer: Batch spectral shrinkage for safe transfer learning. NeurIPS (2019)

    2019

  8. [8]

    In: ICML (2025)

    Cheng, R., Xiong, F., Wei, Y., Zhu, W., Yuan, C.: Whoever started the interference should end it: Guiding data-free model merging via task vectors. In: ICML (2025)

    2025

  9. [9]

    In: ICML (2026)

    Choi, H., Ahn, D., Lee, Y., Kang, T., Cho, S., Choi, J.: Scale: Self-uncertainty conditioned adaptive looking and execution for vision-language-action models. In: ICML (2026)

    2026

  10. [10]

    In: EMNLP Workshop (2024)

    Chronopoulou, A., Pfeiffer, J., Maynez, J., Wang, X., Ruder, S., Agrawal, P.: Language and task arithmetic with parameter-efficient layers for zero-shot summa- rization. In: EMNLP Workshop (2024)

    2024

  11. [11]

    In: CoRL Workshop (2025)

    Dass, S., Khaddaj, A., Engstrom, L., Madry, A., Ilyas, A., Martín-Martín, R.: Datamil: Selecting data for robot imitation learning with datamodels. In: CoRL Workshop (2025)

    2025

  12. [12]

    In: ICRA (2025)

    Dey, S., Zaech, J.N., Nikolov, N., Van Gool, L., Paudel, D.P.: Revla: Reverting visual domain limitation of robotic foundation models. In: ICRA (2025)

    2025

  13. [13]

    In: CVPR (2026)

    Fei, S., Wang, S., Shi, J., Dai, Z., Cai, J., Qian, P., Ji, L., He, X., Zhang, S., Fei, Z., et al.: Libero-plus: In-depth robustness analysis of vision-language-action models. In: CVPR (2026)

    2026

  14. [14]

    In: CVPR (2026)

    Fu, Y., Zhang, Z., Zhang, Y., Wang, Z., Huang, Z., Luo, Y.: Mergevla: Cross-skill model merging toward a generalist vision-language-action agent. In: CVPR (2026)

    2026

  15. [15]

    RA-L (2026)

    Gao, J., Belkhale, S., Dasari, S., Balakrishna, A., Shah, D., Sadigh, D.: A taxonomy for evaluating generalist robot manipulation policies. RA-L (2026)

    2026

  16. [16]

    In: CVPR (2025)

    Gargiulo, A.A., Crisostomi, D., Bucarelli, M.S., Scardapane, S., Silvestri, F., Rodola, E.: Task singular vectors: Reducing task interference in model merging. In: CVPR (2025)

    2025

  17. [17]

    SIAM review53(2), 217–288 (2011)

    Halko, N., Martinsson, P.G., Tropp, J.A.: Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions. SIAM review53(2), 217–288 (2011)

    2011

  18. [18]

    In: ICLR (2022) Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts 17

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: ICLR (2022) Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts 17

    2022

  19. [19]

    In: ACL (2024)

    Huang, S.C., Li, P.Z., Hsu, Y.C., Chen, K.M., Lin, Y.T., Hsiao, S.K., Tsai, R., Lee, H.Y.: Chat vector: A simple approach to equip llms with instruction following and model alignment in new languages. In: ACL (2024)

    2024

  20. [20]

    In: ICLR (2023)

    Ilharco, G., Ribeiro, M.T., Wortsman, M., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. In: ICLR (2023)

    2023

  21. [21]

    $\pi^{*}_{0.6}$: a VLA That Learns From Experience

    Intelligence, P., Amin, A., Aniceto, R., Balakrishna, A., Black, K., Conley, K., Connors, G., Darpinian, J., Dhabalia, K., DiCarlo, J., et al.:π∗ 0.6: a vla that learns from experience. arXiv preprint arXiv:2511.14759 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    In: CoRL (2024)

    Iyer, A., Peng, Z., Dai, Y., Guzey, I., Haldar, S., Chintala, S., Pinto, L.: Open teach: A versatile teleoperation system for robotic manipulation. In: CoRL (2024)

    2024

  23. [23]

    In: ECCV (2024)

    Jang, D.H., Yun, S., Han, D.: Model stock: All we need is just a few fine-tuned models. In: ECCV (2024)

    2024

  24. [24]

    In: ICLR (2025)

    Jin, R., Hou, B., Xiao, J., Su, W.J., Shen, L.: Fine-tuning attention modules only: Enhancing weight disentanglement in task arithmetic. In: ICLR (2025)

    2025

  25. [25]

    In: RSS (2024)

    Khazatsky, A., Pertsch, K., Nair, S., Balakrishna, A., Dasari, S., Karamcheti, S., Nasiriany, S., Srirama, M.K., Chen, L.Y., Ellis, K., et al.: Droid: A large-scale in-the-wild robot manipulation dataset. In: RSS (2024)

    2024

  26. [26]

    In: RSS (2025)

    Kim, M.J., Finn, C., Liang, P.: Fine-tuning vision-language-action models: Opti- mizing speed and success. In: RSS (2025)

    2025

  27. [27]

    In: CoRL (2024)

    Kim, M.J., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., Rafailov, R., Foster, E.P., Sanketi, P.R., Vuong, Q., et al.: Openvla: An open-source vision- language-action model. In: CoRL (2024)

    2024

  28. [28]

    In: CoRL (2025)

    Kumar, S., Dass, S., Pavlakos, G., Martín-Martín, R.: Collage: Adaptive fusion- based retrieval for augmented policy learning. In: CoRL (2025)

    2025

  29. [29]

    In: ICRA (2024)

    Lawson, D., Qureshi, A.H.: Merging decision transformers: Weight averaging for forming multi-task policies. In: ICRA (2024)

    2024

  30. [30]

    In: ICML (2024)

    Lee, S., Wang, Y., Etukuru, H., Kim, H.J., Shafiullah, N.M.M., Pinto, L.: Behavior generation with latent actions. In: ICML (2024)

    2024

  31. [31]

    In: ICML (2024)

    Li, N., Pan, A., Gopal, A., Yue, S., Berrios, D., Gatti, A., Li, J.D., Dombrowski, A.K., Goel, S., Mukobi, G., Helm-Burger, N., Lababidi, R., Justen, L., Liu, A.B., Chen, M., Barrass, I., Zhang, O., Zhu, X., Tamirisa, R., Bharathi, B., Herbert-Voss, A., Breuer, C.B., Zou, A., Mazeika, M., Wang, Z., Oswal, P., Lin, W., Hunt, A.A., Tienken-Harder, J., Shih,...

    2024

  32. [32]

    In: CVPR (2026)

    Li, W., Zhang, Q., Zhai, R., Lin, L., Wang, G.: Vla models are more generalizable than you think: Revisiting physical and spatial modeling. In: CVPR (2026)

    2026

  33. [33]

    In: CoRL (2024)

    Li, X., Hsu, K., Gu, J., Mees, O., Pertsch, K., Walke, H.R., Fu, C., Lunawat, I., Sieh, I., Kirmani, S., et al.: Evaluating real-world robot manipulation policies in simulation. In: CoRL (2024)

    2024

  34. [34]

    In: ICML (2026)

    Li, Y., Peng, Z., Zhang, J., Guo, J., Duan, Y., Shi, Y.: When shared knowledge hurts: Spectral over-accumulation in model merging. In: ICML (2026)

    2026

  35. [35]

    arXiv preprint arXiv:2507.00416 (2025)

    Lin, T., Li, G., Zhong, Y., Zou, Y., Du, Y., Liu, J., Gu, E., Zhao, B.: Evo-0: Vision-language-action model with implicit spatial understanding. arXiv preprint arXiv:2507.00416 (2025)

    work page arXiv 2025

  36. [36]

    In: ICLR (2023)

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023)

    2023

  37. [37]

    In: NeurIPS (2023) 18 T

    Liu, B., Zhu, Y., Gao, C., Feng, Y., Liu, Q., Zhu, Y., Stone, P.: Libero: Benchmarking knowledge transfer for lifelong robot learning. In: NeurIPS (2023) 18 T. Kang et al

    2023

  38. [38]

    In: NeurIPS (2023)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)

    2023

  39. [39]

    In: CVPR (2026)

    Liu, S., Yin, Y., Wang, L., Fan, Q., Shi, Y., Li, W., Gao, Y., Tao, D.: Understanding and enforcing weight disentanglement in task arithmetic. In: CVPR (2026)

    2026

  40. [40]

    In: CVPRW (2026)

    Liu, S., Singh, I.S., Xu, Y., Duan, J., Krishna, R.: Vls: Steering pretrained robot policies via vision-language models. In: CVPRW (2026)

    2026

  41. [41]

    In: ICLR (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)

    2019

  42. [42]

    In: CoRL (2023)

    Mandlekar, A., Nasiriany, S., Wen, B., Akinola, I., Narang, Y., Fan, L., Zhu, Y., Fox, D.: Mimicgen: A data generation system for scalable robot learning using human demonstrations. In: CoRL (2023)

    2023

  43. [43]

    In: ICML (2025)

    Marczak, D., Magistri, S., Cygert, S., Twardowski, B., Bagdanov, A.D., van de Weijer, J.: No task left behind: Isotropic model merging with common and task- specific subspaces. In: ICML (2025)

    2025

  44. [44]

    In: NeurIPS (2022)

    Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in gpt. In: NeurIPS (2022)

    2022

  45. [45]

    In: ICLR (2023)

    Meng, K., Sen Sharma, A., Andonian, A., Belinkov, Y., Bau, D.: Mass editing memory in a transformer. In: ICLR (2023)

    2023

  46. [46]

    In: CoRL (2022)

    Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. In: CoRL (2022)

    2022

  47. [47]

    In: NeurIPS (2023)

    Ortiz-Jimenez, G., Favero, A., Frossard, P.: Task arithmetic in the tangent space: Improved editing of pre-trained models. In: NeurIPS (2023)

    2023

  48. [48]

    In: ICRA (2024)

    O’Neill, A., Rehman, A., Maddukuri, A., Gupta, A., Padalkar, A., Lee, A., Pooley, A., Gupta, A., Mandlekar, A., Jain, A., et al.: Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In: ICRA (2024)

    2024

  49. [49]

    In: NeurIPS (2025)

    Panariello, A., Marczak, D., Magistri, S., Porrello, A., Twardowski, B., Bagdanov, A.D., Calderara, S., van de Weijer, J.: Accurate and efficient low-rank model merging in core space. In: NeurIPS (2025)

    2025

  50. [50]

    In: RSS (2025)

    Pertsch, K., Stachowicz, K., Ichter, B., Driess, D., Nair, S., Vuong, Q., Mees, O., Finn, C., Levine, S.: Fast: Efficient action tokenization for vision-language-action models. In: RSS (2025)

    2025

  51. [51]

    In: EMNLP (2025)

    Qiu, H., Wu, Y., Li, D., Guo, J., Yao, Q.: Superpose task-specific features for model merging. In: EMNLP (2025)

    2025

  52. [52]

    arXiv preprint arXiv:2508.21112 (2025)

    Qu, D., Song, H., Chen, Q., Chen, Z., Gao, X., Ye, X., Lv, Q., Shi, M., Ren, G., Ruan, C., et al.: Eo-1: Interleaved vision-text-action pretraining for general robot control. arXiv preprint arXiv:2508.21112 (2025)

    work page arXiv 2025

  53. [53]

    In: ICLR (2026)

    Seo, M., Kim, T., Lee, H., Choi, J., Tuytelaars, T.: Not all clients are equal: Collaborative model personalization on heterogeneous multi-modal clients. In: ICLR (2026)

    2026

  54. [54]

    In: NeurIPS (2022)

    Shafiullah, N.M., Cui, Z., Altanzaya, A.A., Pinto, L.: Behavior transformers: Cloning kmodes with one stone. In: NeurIPS (2022)

    2022

  55. [55]

    Action Hallucination in Generative Vision-Language-Action Models

    Soh, H., Lim, E.: Action hallucination in generative visual-language-action models. arXiv preprint arXiv:2602.06339 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  56. [56]

    In: ICLR (2025)

    Stoica, G., Ramesh, P., Ecsedi, B., Choshen, L., Hoffman, J.: Model merging with svd to tie the knots. In: ICLR (2025)

    2025

  57. [57]

    In: ICML (2025)

    Sun, W., Li, Q., Geng, Y., Li, B.: Cat merging: A training-free approach for resolving conflicts in model merging. In: ICML (2025)

    2025

  58. [58]

    In: ACL (2025)

    Thakkar, M., Fournier, Q., Riemer, M., Chen, P.Y., Zouaq, A., Das, P., Chandar, S.: Combining domain and alignment vectors to achieve better knowledge-safety trade-offs in llms. In: ACL (2025)

    2025

  59. [59]

    In: IROS (2012) Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts 19

    Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: IROS (2012) Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts 19

    2012

  60. [60]

    In: CoRL (2023)

    Walke, H.R., Black, K., Zhao, T.Z., Vuong, Q., Zheng, C., Hansen-Estruch, P., He, A.W., Myers, V., Kim, M.J., Du, M., et al.: Bridgedata v2: A dataset for robot learning at scale. In: CoRL (2023)

    2023

  61. [61]

    In: ICLR (2024)

    Wang, L., Zhang, K., Zhou, A., Simchowitz, M., Tedrake, R.: Robot fleet learning via policy merging. In: ICLR (2024)

    2024

  62. [62]

    In: ICLR (2026)

    Wei, Y., Cheng, R., Jin, W., Yang, E., Shen, L., Hou, L., Du, S., Yuan, C., Cao, X., Tao, D.: Optmerge: Unifying multimodal llm capabilities and modalities via model merging. In: ICLR (2026)

    2026

  63. [63]

    In: CoRL (2025)

    Wilcox, A., Ghanem, M., Moghani, M., Barroso, P., Joffe, B., Garg, A.: Adapt3r: Adaptive 3d scene representation for domain transfer in imitation learning. In: CoRL (2025)

    2025

  64. [64]

    In: CoRL (2025)

    Xie, A., Chand, R., Sadigh, D., Hejna, J.: Data retrieval with importance weights for few-shot imitation learning. In: CoRL (2025)

    2025

  65. [65]

    In: ICRA (2024)

    Xie, A., Lee, L., Xiao, T., Finn, C.: Decomposing the generalization gap in imitation learning for visual robotic manipulation. In: ICRA (2024)

    2024

  66. [66]

    In: NeurIPS (2023)

    Yadav, P., Tam, D., Choshen, L., Raffel, C.A., Bansal, M.: Ties-merging: Resolving interference when merging models. In: NeurIPS (2023)

    2023

  67. [67]

    In: ICLR (2026)

    Yadav, Y., Zhou, Z., Wagenmaker, A., Pertsch, K., Levine, S.: Robust finetuning of vision-language-action robot policies via parameter merging. In: ICLR (2026)

    2026

  68. [68]

    In: NeurIPS (2025)

    Yang, J., Jin, D., Tang, A., Shen, L., Zhu, D., Chen, Z., Zhao, Z., Wang, D., Cui, Q., Zhang, Z., et al.: Mix data or merge models? balancing the helpfulness, honesty, and harmlessness of large language model via model merging. In: NeurIPS (2025)

    2025

  69. [69]

    In: RSS (2025)

    Yang, S., Yu, W., Zeng, J., Lv, J., Ren, K., Lu, C., Lin, D., Pang, J.: Novel demon- stration generation with gaussian splatting enables robust one-shot manipulation. In: RSS (2025)

    2025

  70. [70]

    arXiv preprint arXiv:2602.09021 (2026)

    Yu, C., Sima, C., Jiang, G., Zhang, H., Mai, H., Li, H., Wang, H., Chen, J., Wu, K., Chen, L., Zhao, L., Shi, M., Luo, P., Bu, Q., Peng, S., Li, T., Yuan, Y.:χ0: Resource-aware robust manipulation via taming distributional inconsistencies. arXiv preprint arXiv:2602.09021 (2026)

    work page arXiv 2026

  71. [71]

    In: ICML (2024)

    Yu, L., Yu, B., Yu, H., Huang, F., Li, Y.: Language models are super mario: Absorbing abilities from homologous models as a free lunch. In: ICML (2024)

    2024

  72. [72]

    Scaling Robot Learning with Semantically Imagined Experience

    Yu, T., Xiao, T., Stone, A., Tompson, J., Brohan, A., Wang, S., Singh, J., Tan, C., Peralta, J., Ichter, B., et al.: Scaling robot learning with semantically imagined experience. In: arXiv preprint arXiv:2302.11550 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  73. [73]

    In: CVPR (2025)

    Yun, S., Chae, S., Lee, D., Ro, Y.: Soma: Singular value decomposed minor com- ponents adaptation for domain generalizable representation learning. In: CVPR (2025)

    2025

  74. [74]

    In: ICRA (2025)

    Zhang, W., Li, Y., Qiao, Y., Huang, S., Liu, J., Dayoub, F., Ma, X., Liu, L.: Effective tuning strategies for generalist robot manipulation policies. In: ICRA (2025)

    2025

  75. [75]

    In: ICML (2024)

    Zhao, J., Zhang, Z., Chen, B., Wang, Z., Anandkumar, A., Tian, Y.: Galore: Memory-efficient llm training by gradient low-rank projection. In: ICML (2024)

    2024

  76. [76]

    In: NAACL (2025)

    Zhao, Y., Zhang, W., Wang, H., Kawaguchi, K., Bing, L.: Adamergex: Cross-lingual transfer with large language models via adaptive adapter merging. In: NAACL (2025)

    2025

  77. [77]

    LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization

    Zhou, X., Xu, Y., Tie, G., Chen, Y., Zhang, G., Chu, D., Zhou, P., Sun, L.: Libero- pro: Towards robust and fair evaluation of vision-language-action models beyond memorization. arXiv preprint arXiv:2510.03827 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  78. [78]

    In: ICML (2024)

    Zhou, Z., Chen, Z., Chen, Y., Zhang, B., Yan, J.: On the emergence of cross-task linearity in the pretraining-finetuning paradigm. In: ICML (2024)

    2024

  79. [79]

    RA-L (2026) 20 T

    Zhu, R., Sun, E., Huang, G., Celiktutan, O.: Efficient continual imitation learning with online meta-adapters. RA-L (2026) 20 T. Kang et al

    2026

  80. [80]

    Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., et al.: Rt-2: Vision-language-action models transfer web knowledge to robotic control. In: CoRL (2023) Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts 21 Supplementary Material This supplementary material provides additional technic...

    2023