Pith. sign in

REVIEW 3 major objections 7 minor 33 references

Unsupervised skill mining turns multi-task robot demos into a reusable prior for few-shot adaptation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 08:49 UTC pith:BBRXS4ZN

load-bearing objection Solid plug-in skill mining with real multi-seed and real-robot gains; the fixed-library freeze is a real but already-acknowledged limit, not a hidden collapse of the claim. the 3 major comments →

arxiv 2607.08354 v1 pith:BBRXS4ZN submitted 2026-07-09 cs.RO

SkillPlug: Unsupervised Skill Mining for Few-Shot Adaptation in Robotic Manipulation

Zi-han Ding , Ziwei Wang This is my paper
classification cs.RO
keywords skill miningfew-shot adaptationvisuomotor imitationrobotic manipulationskill librarycompositional controlself-supervised skills
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

End-to-end visuomotor imitation policies map observations straight to low-level actions, so shared behavioral structure stays buried and transfer to new tasks needs many demos. SkillPlug plugs a skill library, interactor, and router into an existing policy and mines that library from raw multi-task demonstrations with self-supervised losses that favor compact, behavior-aligned, non-redundant primitives. After mining, the skills stay frozen; only a lightweight router and action head are fine-tuned on a handful of new-task demos. Across two simulation suites and a real arm, the fixed library raises both multi-task success and few-shot success, showing that an explicit reusable skill prior can make imitation learning more compositional and data-efficient without full retraining.

Core claim

A shared library of skills mined unsupervised from multi-task demonstrations, once frozen, supplies a transferable behavioral prior that improves both multi-task performance and few-shot adaptation when only the router and action head are specialized. Gains appear consistently for compact and large backbones in simulation and on a real robot.

What carries the argument

SkillPlug: a plug-in skill library of continuous embeddings, a trajectory-only VAE-style posterior used only in training, a cross-attention interactor that conditions scene features on each skill, and a router that mixes skill-conditioned features; trained by reconstruction, KL compactness, behavioral skill alignment, and skill-disentanglement losses, then frozen for adaptation.

Load-bearing premise

Skills mined once from the training task distribution stay sufficient for new tasks when they are frozen and only the router and action head can change; if a new task needs a primitive outside that library, adaptation cannot invent it.

What would settle it

Take a novel manipulation task whose required motion primitive is absent from the multi-task training set; after freezing the mined skills and fine-tuning only router and action head on a few demos, success should remain near the no-skill baseline rather than rise.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. SkillPlug is a plug-in, architecture-agnostic module that augments an existing visuomotor imitation policy with a shared continuous skill library, a skill interactor, and a chunk-level router. Skills are mined unsupervised from multi-task demonstrations via a trajectory-only VAE-style posterior and four self-supervised losses (reconstruction, KL compactness, Behavioral Skill Alignment, and Skill Disentanglement). After multi-task training the skill embeddings are frozen; few-shot specialization updates only the router and action head. On DISCOVERSE (ACT) and LIBERO (OpenVLA-OFT), and on a real Galaxea arm, the method reports large multi-task and few-shot gains relative to the corresponding vanilla baselines, with supporting ablations and qualitative skill analyses.

Significance. If the reported gains hold under fairer controls, SkillPlug is a practically useful contribution: it is backbone-agnostic, requires no skill labels, and converts multi-task data into a reusable behavioral prior that improves data-efficient adaptation. Strengths include multi-seed evaluation on two simulation benchmarks spanning compact and large VLA backbones, a progressive loss-term ablation, skill-count and redundancy analysis, inference-step efficiency measurements, and real-robot multi-task/few-shot trials. The train-then-freeze protocol and explicit non-redundancy objectives are concrete design choices that the community can reuse. The work is a solid fit for RA-L-style robotics methods papers.

major comments (3)
  1. Sec. IV-A few-shot protocol: SkillPlug fine-tunes router + action head while the vanilla baseline fine-tunes only the action head. This confounds attribution of the large few-shot gains (Table I: +18.1 pts average; Table II: +38.3 pts average; real robot +28.5 pts) to the mined skill prior versus extra tunable capacity and modular structure. A load-bearing control is needed: e.g., (i) SkillPlug with frozen skills and frozen router (action head only), (ii) SkillPlug with frozen skills but trainable router, and (iii) a capacity-matched vanilla baseline that adds a lightweight adapter of similar parameter count. Without this, the central claim that frozen mined skills are the transferable prior is only partially supported for the few-shot setting (multi-task comparisons remain fair).
  2. Sec. II-B positions SkillPlug against data-driven skill methods (trajectory clustering, learnable skill embeddings, MoE experts as skills), but Sec. IV only compares to end-to-end ACT / OpenVLA-OFT. At least one strong skill/MoE baseline under the same multi-task then few-shot protocol is needed to show that the proposed BSA/SD objectives and trajectory-only posterior, not merely “adding a skill/router layer,” drive the gains. This is load-bearing for the paper’s claim of unsupervised skill mining as the contribution rather than hierarchical conditioning alone.
  3. Sec. III-B train-then-freeze and Sec. V: the strongest claim assumes the multi-task-mined library is a sufficient compositional prior for unseen tasks. All reported few-shot tasks (DISCOVERSE cuplid/block/jujube/mouse; LIBERO cross-suite; real stand-cup / banana-bread) appear to reuse low-level primitives already present in the multi-task distribution. The conclusions correctly flag the fixed-library limitation, but the manuscript should either (a) quantify primitive overlap / failure cases when a required motion is outside the library, or (b) add a stress-test task that needs a clearly novel primitive. Without that, the scope of “few-shot adaptation to unseen tasks” remains underspecified relative to the claim.
minor comments (7)
  1. Abstract and intro cite “+45.1% and +18.1% points” on DISCOVERSE; Table I reports per-task means but not the exact aggregate used. State the aggregation (macro-average over tasks/seeds) explicitly next to the headline numbers.
  2. Eq. (7)–(11): the mutual-information motivation for BSA is informal (I(τ;s) is not estimated). A short note that BSA is a practical soft-label alignment surrogate, not a tight MI estimator, would avoid overclaiming.
  3. Table III: ablated rows are single-seed while the full objective is 3-seed ±SE. Either run ablations with multiple seeds or mark more clearly that only the full row is multi-seed.
  4. Figs. 3–4 and 5–6: layout and repeated “open gripper / move forward” panels make it easy to confuse scene-agnosticity with compositionality. Separate captions and a one-line takeaway under each figure would help.
  5. Hyperparameters: report λ_KL, λ_BSA, λ_SD, d_s, and interactor architecture (layers/heads) in the main text or a short appendix table; currently only K and high-level training settings are given.
  6. Notation: τ_t is used both for the predicted action chunk (Eq. 1) and as a trajectory segment in the posterior/BSA; a brief disambiguation would reduce confusion.
  7. Real-robot Fig. 7 reports successes out of 20 without error bars or seeds; even 2–3 seeds or binomial CIs would strengthen the claim.

Circularity Check

0 steps flagged

No circularity: empirical skill-mining method with held-out multi-task and few-shot evaluation; objectives do not define the reported success metrics by construction.

full rationale

SkillPlug is an empirical imitation-learning methods paper, not a first-principles derivation. The skill library is optimized with reconstruction (behavior cloning), KL, BSA, and SD losses on multi-task demonstration data; reported claims are success rates on benchmark-defined held-out trials and on few-shot tasks unseen during multi-task mining. BSA explicitly uses stop-gradient on router soft targets so skill embeddings are shaped without co-defining the evaluation metric with the router. Train-then-freeze (fixed skills, tune router/action head) is a design protocol whose sufficiency is an empirical assumption, not a tautology: gains are measured against vanilla baselines under the same few-shot budget, not forced by fitting the reported SR numbers. There is no self-definitional loop, no fitted constant renamed as a prediction, no load-bearing uniqueness theorem from overlapping authors, and no ansatz smuggled in via self-citation. Related-work citations (ACT, OpenVLA-OFT, LIBERO, etc.) are external baselines and datasets. Minor hyperparameter choices (K, λ weights) are ordinary method design, not circular reductions. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 5 invented entities

The central claim rests on standard imitation-learning assumptions plus paper-specific design choices: a fixed discrete skill library, action-only skill mining, and train-then-freeze adaptation. Free parameters (K, loss weights, embedding sizes) and invented modules (skill embeddings, posterior encoder, BSA/SD losses, interactor/router) carry most of the method's novelty; independent evidence for the skill entities is only the paper's own ablations and visualizations, not external measurements.

free parameters (4)
  • number of skills K
    Chosen as K=4 for ACT/DISCOVERSE and K=8 for OpenVLA-OFT/LIBERO; Table IV shows performance depends on K and saturates with rising redundancy.
  • loss weights λ_KL, λ_BSA, λ_SD
    Eq. (14) is a weighted sum; exact weights are not reported as fixed by theory and must be set by the authors for the reported runs.
  • skill embedding dimension d_s and interactor capacity
    Continuous skill codes and cross-attention interactor size are architectural free choices that define the skill space capacity.
  • action chunk length / prediction horizon n
    Chunk size 25 on DISCOVERSE and OpenVLA-OFT defaults on LIBERO structure what counts as a skill-aligned trajectory segment.
axioms (4)
  • domain assumption Offline multi-task demonstration datasets contain reusable, composable behavior-level primitives that can be mined without skill labels.
    Stated throughout Sec. I and III-A as the premise for unsupervised skill mining from raw demos.
  • ad hoc to paper Conditioning a trajectory-skill posterior only on actions (not observations) yields scene-agnostic behavioral skills.
    Sec. III-B introduces the action-only VAE-style posterior as an information bottleneck; this is a design axiom, not a proved fact.
  • ad hoc to paper After multi-task mining, freezing skill embeddings and fine-tuning only router and action head is sufficient for few-shot specialization.
    Train-then-freeze protocol in Sec. III-B/IV-A is load-bearing for the few-shot claim; conclusions admit failure when new primitives are required.
  • domain assumption Behavior cloning with L1 action-chunk loss is a valid primary supervision signal for visuomotor policies.
    Standard imitation-learning assumption used in Eq. (5) and throughout experiments.
invented entities (5)
  • Shared continuous skill library {s_k} no independent evidence
    purpose: Explicit reusable skill codes shared across tasks and frozen at adaptation time.
    Postulated as learnable embeddings that represent behavior primitives; evidence is internal success rates and qualitative probes only.
  • Trajectory–skill posterior encoder q_φ no independent evidence
    purpose: Action-only bottleneck to bias skills toward trajectory structure rather than scene appearance.
    Training-only VAE-style module discarded at inference; no external validation beyond ablations.
  • Behavioral Skill Alignment (BSA) loss no independent evidence
    purpose: Align skill embeddings with trajectory-level behavior via soft-label classification against router targets.
    Paper-specific objective (Eqs. 7–11) motivated by mutual information / InfoNCE-style scoring.
  • Skill Disentanglement (SD) loss no independent evidence
    purpose: Penalize similarity of skill-conditioned features under the same observation to reduce redundancy.
    Paper-specific regularizer (Eqs. 12–13); supported only by internal ablation gains.
  • Skill interactor + chunk-level router no independent evidence
    purpose: Produce and mix skill-conditioned features for the frozen action head.
    Architectural invention enabling plug-in conditioning without changing backbone I/O shapes.

pith-pipeline@v1.1.0-grok45 · 18287 in / 3767 out tokens · 40310 ms · 2026-07-10T08:49:43.742097+00:00 · methodology

0 comments
read the original abstract

Learning transferable visuomotor imitation policies that generalize across diverse manipulation tasks and adapt rapidly to new tasks from only a handful of demonstrations remains challenging. Most modern policies are trained end-to-end to map observations directly to low-level actions, offering little explicit structure for reusing and recombining behaviors across tasks and making transfer data-inefficient under limited supervision. We propose SkillPlug, a plug-in framework that augments an existing visuomotor policy with a skill-conditioning module and mines a shared, transferable skill library from raw multi-task demonstrations. SkillPlug learns skills via self-supervised objectives that promote compact, reusable, and non-redundant behavior-level primitives, forming a task-shared prior for compositional control. After skill mining, we keep the learned skills fixed and specialize to unseen tasks by fine-tuning only lightweight router and action head, enabling efficient adaptation without full end-to-end retraining. We evaluate SkillPlug on two simulation benchmarks and on a real robot, and observe that the mined transferable skills consistently improve both multi-task performance and few-shot adaptation. Overall, SkillPlug offers a scalable way to mine reusable skills that improve data-efficient generalization in robotic manipulation.

Figures

Figures reproduced from arXiv: 2607.08354 by Zi-Han Ding, Ziwei Wang.

Figure 1
Figure 1. Figure 1: Conventional imitation-based visuomotor policies often struggle [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SkillPlug architecture and training. SkillPlug augments a base visuomotor policy with a skill library, a skill interactor, and a router that composes [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Scene-agnostic behavior of a single skill. The selected skill consis [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Compositionality via routing. We hard-code the router to select [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Interpretable routing dynamics. Top: key frames of a [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Real-robot performance of ACT and ACT+SkillPlug. Left: multi-task [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Skill diversity and routing dynamics. (a) Cosine-similarity matrices [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Skill routing visualization on a real robot for [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗

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

33 extracted references · 33 canonical work pages · 8 internal anchors

  1. [1]

    Bridgedata v2: A dataset for robot learning at scale,

    H. R. Walke, K. Black, T. Z. Zhao, Q. Vuong, C. Zheng, P. Hansen- Estruch, A. W. He, V . Myers, M. J. Kim, M. Duet al., “Bridgedata v2: A dataset for robot learning at scale,” inConference on Robot Learning. PMLR, 2023, pp. 1723–1736

    work page 2023

  2. [2]

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collab- oration,

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jainet al., “Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collab- oration,” inICRA, 2024. 8 IEEE ROBOTICS AND AUTOMATION LETTERS. PREPRINT VERSION. ACCEPTED JUNE, 2026 Encoder 𝜇𝜎2 Decoder Action Head Scene-Context En...

    work page 2024

  3. [3]

    Libero: Benchmarking knowledge transfer for lifelong robot learning,

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “Libero: Benchmarking knowledge transfer for lifelong robot learning,”Advances in Neural Information Processing Systems, vol. 36, pp. 44 776–44 791, 2023

    work page 2023

  4. [4]

    RT-1: Robotics Transformer for Real-World Control at Scale

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsuet al., “Rt-1: Robotics transformer for real-world control at scale,”arXiv preprint arXiv:2212.06817, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  5. [5]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahidet al., “Rt-2: Vision-language-action models transfer web knowledge to robotic control,” inConference on Robot Learning. PMLR, 2023, pp. 2165–2183

    work page 2023

  6. [6]

    Openvla: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuonget al., “Openvla: An open-source vision-language-action model,” inConference on Robot Learning. PMLR, 2025, pp. 2679–2713

    work page 2025

  7. [7]

    Octo: An open-source generalist robot policy,

    O. Mees, D. Ghosh, K. Pertsch, K. Black, H. R. Walke, S. Dasari, J. Hejna, T. Kreiman, C. Xu, J. Luoet al., “Octo: An open-source generalist robot policy,” inFirst Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024, 2024

    work page 2024

  8. [8]

    π: A vision-language-action flow model for general robot control,

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichteret al., “π: A vision-language-action flow model for general robot control,”CoRR, 2024

    work page 2024

  9. [9]

    $\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusaiet al., “π0. 5: A vision-language- action model with open-world generalization. arxiv 2025,”arXiv preprint arXiv:2504.16054

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Cronusvla: Transferring latent motion across time for multi-frame prediction in manipulation,

    H. Li, S. Yang, Y . Chen, Y . Tian, X. Yang, X. Chen, H. Wang, T. Wang, F. Zhao, D. Linet al., “Cronusvla: Transferring latent motion across time for multi-frame prediction in manipulation,”arXiv preprint arXiv:2506.19816, 2025

    work page arXiv 2025

  11. [11]

    Query-centric diffusion policy for generalizable robotic assembly,

    Z. Xu, H. Lin, S. Liu, and D. Zhao, “Query-centric diffusion policy for generalizable robotic assembly,”arXiv preprint arXiv:2509.18686, 2025

    work page arXiv 2025

  12. [12]

    Few-shot neuro- symbolic imitation learning for long-horizon planning and acting,

    P. Lorang, H. Lu, J. Huemer, P. Zips, and M. Scheutz, “Few-shot neuro- symbolic imitation learning for long-horizon planning and acting,” in Conference on Robot Learning. PMLR, 2025, pp. 2501–2518

    work page 2025

  13. [13]

    Long-vla: Unleashing long-horizon capability of vision language action model for robot manipulation,

    Y . Fan, S. Bai, X. Tong, P. Ding, Y . Zhu, H. Lu, F. Dai, W. Zhao, Y . Liu, S. Huanget al., “Long-vla: Unleashing long-horizon capability of vision language action model for robot manipulation,” inConference on Robot Learning. PMLR, 2025, pp. 2018–2037

    work page 2025

  14. [14]

    Compose by focus: Scene graph-based atomic skills,

    H. Qi, C. Chen, and H. Yang, “Compose by focus: Scene graph-based atomic skills,”arXiv preprint arXiv:2509.16053, 2025

    work page arXiv 2025

  15. [15]

    Generalizable hierarchical skill learning via object-centric representation,

    H. Zhao, Y . Qi, B. Hu, Y . Zhu, Z. Chen, H. Tian, X. Zhu, O. Howell, H. Huang, R. Walterset al., “Generalizable hierarchical skill learning via object-centric representation,”arXiv preprint arXiv:2510.21121, 2025

    work page arXiv 2025

  16. [16]

    Language-guided skill learning with temporal vari- ational inference,

    H. Fu, P. Sharma, E. Stengel-Eskin, G. Konidaris, N. Le Roux, M.-A. Cˆot´e, and X. Yuan, “Language-guided skill learning with temporal vari- ational inference,” inProceedings of the 41st International Conference on Machine Learning, 2024, pp. 14 135–14 156

    work page 2024

  17. [17]

    From Experts to a Generalist: Toward General Whole-Body Control for Humanoid Robots

    Y . Wang, M. Yang, Z. Ding, Y . Zhang, W. Zeng, X. Xu, H. Jiang, and Z. Lu, “From experts to a generalist: Toward general whole-body control for humanoid robots,”arXiv preprint arXiv:2506.12779, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    Universal actions for enhanced embodied foundation models,

    J. Zheng, J. Li, D. Liu, Y . Zheng, Z. Wang, Z. Ou, Y . Liu, J. Liu, Y .-Q. Zhang, and X. Zhan, “Universal actions for enhanced embodied foundation models,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 22 508–22 519

    work page 2025

  19. [19]

    Discrete latent plans via semantic skill abstractions

    H. Jiang, J. Wang, and Z. Lu, “Discrete latent plans via semantic skill abstractions.” inICLR, 2025

    work page 2025

  20. [20]

    Semantic temporal abstraction via vision-language model guidance for efficient reinforcement learning,

    T.-S. Liu, X.-H. Liu, R. Chen, L. Jin, P. Wang, Z. Zhang, and Y . Yu, “Semantic temporal abstraction via vision-language model guidance for efficient reinforcement learning,” inThe Thirteenth International Conference on Learning Representations

    work page

  21. [21]

    Think small, act big: Primitive prompt learning for lifelong robot manipulation,

    Y . Yao, S. Liu, H. Song, D. Qu, Q. Chen, Y . Ding, B. Zhao, Z. Wang, X. Li, and D. Wang, “Think small, act big: Primitive prompt learning for lifelong robot manipulation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 22 573–22 583

    work page 2025

  22. [22]

    Quest: Self- supervised skill abstractions for learning continuous control,

    A. Mete, H. Xue, A. Wilcox, Y . Chen, and A. Garg, “Quest: Self- supervised skill abstractions for learning continuous control,”Advances in Neural Information Processing Systems, vol. 37, pp. 4062–4089, 2024

    work page 2024

  23. [23]

    SPECI: Skill Prompts based Hierarchical Continual Imitation Learning for Robot Manipulation

    J. Xu and X. Nie, “Speci: Skill prompts based hierarchical con- tinual imitation learning for robot manipulation,”arXiv preprint arXiv:2504.15561, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Moe-dp: An moe-enhanced diffusion policy for robust long-horizon robotic manipulation with skill decomposition and failure recovery,

    B. Cheng, T. Liang, S. Huang, M. Shao, F. Zhang, B. Xu, Z. Xue, and H. Xu, “Moe-dp: An moe-enhanced diffusion policy for robust long-horizon robotic manipulation with skill decomposition and failure recovery,”arXiv preprint arXiv:2511.05007, 2025

    work page arXiv 2025

  25. [25]

    Learning fine-grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,” inICML Workshop on New Frontiers in Learning, Control, and Dynamical Systems

    work page

  26. [26]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,”The International Journal of Robotics Research, vol. 44, no. 10-11, pp. 1684–1704, 2025

    work page 2025

  27. [27]

    Towards more generalizable one-shot visual imitation learning,

    Z. Mandi, F. Liu, K. Lee, and P. Abbeel, “Towards more generalizable one-shot visual imitation learning,” in2022 International conference on robotics and automation (ICRA). IEEE, 2022, pp. 2434–2444

    work page 2022

  28. [28]

    Rdt-1b: a diffusion foundation model for bimanual ma- nipulation,

    S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu, “Rdt-1b: a diffusion foundation model for bimanual ma- nipulation,” inThe Thirteenth International Conference on Learning Representations

    work page

  29. [29]

    LLaMA: Open and Efficient Foundation Language Models

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azharet al., “Llama: Open and efficient foundation language models,”arXiv preprint arXiv:2302.13971, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  30. [30]

    Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou, “Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond,”arXiv preprint arXiv:2308.12966, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  31. [31]

    Paligemma: A versatile 3b vlm for transfer,

    L. Beyer, A. Steiner, A. S. Pinto, A. Kolesnikov, X. Wang, D. Salz, M. Neumann, I. Alabdulmohsin, M. Tschannen, E. Bugliarelloet al., “Paligemma: A versatile 3b vlm for transfer,”CoRR, 2024

    work page 2024

  32. [32]

    DISCOVERSE: Efficient Robot Simulation in Complex High-Fidelity Environments

    Y . Jia, G. Wang, Y . Dong, J. Wu, Y . Zeng, H. Lin, Z. Wang, H. Ge, W. Gu, K. Dinget al., “Discoverse: Efficient robot simulation in complex high-fidelity environments,”arXiv preprint arXiv:2507.21981, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success

    M. J. Kim, C. Finn, and P. Liang, “Fine-tuning vision-language- action models: Optimizing speed and success, 2025,”URL https://arxiv. org/abs/2502.19645

    work page internal anchor Pith review Pith/arXiv arXiv 2025