pith. machine review for the scientific record. sign in

arxiv: 2605.07381 · v1 · submitted 2026-05-08 · 💻 cs.RO · cs.AI

Recognition: no theorem link

Escaping the Diversity Trap in Robotic Manipulation via Anchor-Centric Adaptation

Yanzhe Chen , Kevin Yuchen Ma , Qi Lv , Yiqi Lin , Zechen Bai , Chen Gao , Mike Zheng Shou
Authors on Pith no claims yet

Pith reviewed 2026-05-11 02:21 UTC · model grok-4.3

classification 💻 cs.RO cs.AI
keywords robotic manipulationvision-language-action modelsdata budget adaptationcoverage density trade-offanchor-centric adaptationpolicy error decompositionembodiment gap
0
0 comments X

The pith

Repeating demonstrations at a few core anchor conditions is optimal for adapting robot policies under a fixed data budget.

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

The paper argues that the common strategy of collecting diverse single demonstrations to maximize coverage backfires because it leaves estimation noise unaddressed. It decomposes policy error into a density-dependent estimation term and a coverage-dependent extrapolation term to show that an interior optimum exists for how many unique conditions to sample. This leads to a method that first collects repeated data at stable anchors to build a reliable policy base before carefully adding boundary cases. A sympathetic reader would care because robot data collection is expensive, so smarter allocation can improve reliability without extra cost. Real-robot tests confirm better success rates than standard diverse sampling.

Core claim

The central claim is that policy error decomposes into non-vanishing estimation error from insufficient density and extrapolation error from limited coverage, yielding an interior optimal number of unique conditions for any fixed budget; Anchor-Centric Adaptation exploits this by first stabilizing via repeated anchor demonstrations and then expanding via teacher-forced mining of high-risk boundaries.

What carries the argument

The Coverage-Density Trade-off, derived from decomposing policy error into estimation (density) and extrapolation (coverage) terms, which identifies the interior optimum for demonstration allocation.

Load-bearing premise

The decomposition of policy error into separate estimation and extrapolation terms accurately captures the dynamics of robotic policy adaptation and yields a meaningful interior optimum rather than an all-diverse boundary solution.

What would settle it

An experiment where increasing the number of unique conditions without repeats always reduces error more than the proposed anchor method would falsify the existence of an interior optimum.

Figures

Figures reproduced from arXiv: 2605.07381 by Chen Gao, Kevin Yuchen Ma, Mike Zheng Shou, Qi Lv, Yanzhe Chen, Yiqi Lin, Zechen Bai.

Figure 1
Figure 1. Figure 1: Illustration of Motivation. Top: Contrasting the “di￾versity trap” of sparse, single-shot sampling against stable anchor￾centric repetition. Middle & Bottom: Inverted-U trend of success versus number of anchors, with 3D visualizations of sample distri￾butions and densities at representative points. these models on specific physical platforms remains a chal￾lenge due to the embodiment mismatch and subtle en… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Anchor-Centric Adaptation (ACA). Stage 1 learns a stable core policy by repeating demonstrations at sparse anchors; Boundary mining screens probe trajectories and selects high-deviation locations; Stage 2 expands boundary competence via a constrained, parameter-efficient residual update in the Action Expert, while keeping the pretrained VLM frozen. repeated demonstrations at each anchor (ni > 1… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of (a) the real-robot experimental setup and (b) the spatial definition of S@1–S@3 metrics. modulate the residual strength using the flow-time embed￾ding so the correction can specialize across different stages of the denoising process. Optimization. We optimize only ϕ on the boundary dataset Dbd using the same flow-matching objective as Stage 1, while keeping θ0 frozen. The residual branch is… view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity to the number of anchors (K). Success rates exhibit a consistent inverted-U trend across different budgets. N = 100 trajectories and the number of anchors at K = 6, as shown in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity to anchor consolidation budget (NA). Region-level success rates under a fixed total budget of N = 100.      &$$%  " % ! &#  !%   ! ! "'  !                         Toy Tiding Table Cleaning Cup Placement Block Stacking Success Rate (%) Tasks 𝝅𝟎.𝟓 𝝅𝟎 + ACA [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance gains across different base VLA models. ACA yields consistent performance increases on four tasks. vs. 55.0% in S@3). This indicates that once the policy skeleton is consolidated, dedicating a minor portion of the budget to targeted Stage 2 mining provides larger marginal gains than further anchor repetition. Performance Gains across Different VLAs. We evaluate the architectural compatibility o… view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of real-robot rollouts across four tasks: from top to bottom, Block Stacking, Table Cleaning, Cup Placement, and Toy Tidying; each row shows the task instruction and corresponding key frames from the rollout video. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
read the original abstract

While Vision-Language-Action (VLA) models offer broad general capabilities, deploying them on specific hardware requires real-world adaptation to bridge the embodiment gap. Since robot demonstrations are costly, this adaptation must often occur under a strict data budget. In this work, we identify a critical diversity trap: the standard heuristic of "maximizing coverage" by collecting diverse, single-shot demonstrations can be self-defeating due to non-vanishing estimation noise. We formalize this phenomenon as a Coverage--Density Trade-off. By decomposing the policy error into estimation (density) and extrapolation (coverage) terms, we characterize an interior optimal allocation of unique conditions for a fixed budget. Guided by this analysis, we propose Anchor-Centric Adaptation (ACA), a two-stage framework that first stabilizes a policy skeleton through repeated demonstrations at core anchors, then selectively expands coverage to high-risk boundaries via teacher-forced error mining and constrained residual updates. Real-robot experiments validate our trade-off framework and demonstrate that ACA significantly improves task reliability and success rates over standard diverse sampling strategies under the same budget.

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 identifies a 'diversity trap' in adapting Vision-Language-Action (VLA) models to specific robot hardware under strict data budgets, where maximizing coverage via single-shot diverse demonstrations can increase estimation noise. It formalizes this as a Coverage-Density Trade-off by decomposing policy error into non-vanishing estimation (density) and extrapolation (coverage) terms, deriving an interior optimal allocation of unique conditions for fixed budget. It proposes Anchor-Centric Adaptation (ACA): a two-stage process that first stabilizes a policy skeleton via repeated demonstrations at core anchors, then selectively expands to high-risk boundaries using teacher-forced error mining and constrained residual updates. Real-robot experiments are claimed to validate the trade-off and show ACA yields higher task reliability and success rates than standard diverse sampling under identical budgets.

Significance. If the decomposition is valid and produces a robust interior optimum rather than a boundary solution, the work supplies a principled, budget-aware strategy for data collection in VLA adaptation that could reduce costly robot demonstrations while improving reliability. The real-robot validation is a concrete strength, as is the attempt to move beyond heuristic diversity maximization; however, the significance hinges on whether the claimed interior optimum survives realistic noise and embodiment shifts.

major comments (2)
  1. [Abstract / Coverage-Density Trade-off formalization] Abstract and formalization of Coverage-Density Trade-off: the interior optimal allocation is derived directly from the paper's additive decomposition of policy error into estimation and extrapolation terms. No first-principles derivation or sensitivity analysis is supplied to demonstrate that the minimum lies strictly inside (0, N) rather than at the all-unique or all-repeated boundary under the loss landscapes of actual VLA adaptation with embodiment noise; if the estimation term decays faster than the extrapolation term grows, the diversity trap disappears and the claimed optimum is an artifact of the modeling assumptions.
  2. [Real-robot experiments] Real-robot experiments (validation section): the abstract asserts significant gains in reliability and success rates, yet no details are provided on the number of trials, statistical tests, exact task suite, embodiment shift magnitude, or whether hyper-parameters and condition selection were chosen post-hoc. Without these, it is impossible to determine whether the reported improvements are attributable to ACA or to implementation specifics, undermining the empirical support for the trade-off framework.
minor comments (2)
  1. [Formalization section] Notation for the error decomposition (estimation vs. extrapolation terms) should be introduced with explicit equations early in the formalization section to avoid ambiguity when the interior optimum is later characterized.
  2. [Abstract] The abstract's phrasing of 'non-vanishing estimation noise' would benefit from a brief parenthetical example of how density affects variance in the VLA policy head.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the formalization of the Coverage-Density Trade-off and the reporting of real-robot experiments. We address each point below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract / Coverage-Density Trade-off formalization] Abstract and formalization of Coverage-Density Trade-off: the interior optimal allocation is derived directly from the paper's additive decomposition of policy error into estimation and extrapolation terms. No first-principles derivation or sensitivity analysis is supplied to demonstrate that the minimum lies strictly inside (0, N) rather than at the all-unique or all-repeated boundary under the loss landscapes of actual VLA adaptation with embodiment noise; if the estimation term decays faster than the extrapolation term grows, the diversity trap disappears and the claimed optimum is an artifact of the modeling assumptions.

    Authors: The decomposition follows standard bias-variance analysis in imitation learning: estimation error scales as O(1/sqrt(R)) with repetitions R per condition due to reduced gradient variance, while extrapolation error grows with the uncovered fraction of the condition space. For fixed budget B = K * R the resulting convex objective has closed-form interior minimizer K* proportional to sqrt(B * lambda_est / lambda_ext) whenever both coefficients are positive. The original submission presented this derivation but omitted explicit sensitivity checks. We will add a new subsection with Monte-Carlo simulations that sweep relative decay rates and embodiment-noise magnitudes to confirm the interior optimum persists under VLA fine-tuning loss surfaces. revision: yes

  2. Referee: [Real-robot experiments] Real-robot experiments (validation section): the abstract asserts significant gains in reliability and success rates, yet no details are provided on the number of trials, statistical tests, exact task suite, embodiment shift magnitude, or whether hyper-parameters and condition selection were chosen post-hoc. Without these, it is impossible to determine whether the reported improvements are attributable to ACA or to implementation specifics, undermining the empirical support for the trade-off framework.

    Authors: We agree that the experimental section requires expanded reporting. The revised manuscript will include: 50 independent rollouts per task per method with mean success rates, standard deviations, and p-values from paired t-tests; a full enumeration of the task suite (pick-and-place, stacking, peg insertion with object and lighting variations); quantitative embodiment-shift metrics (joint-torque mismatch of 12-18 % and camera-calibration offsets); and an explicit statement that anchor selection and all hyperparameters were fixed on a held-out validation split prior to final evaluation. These additions will allow readers to evaluate attribution to ACA. revision: yes

Circularity Check

0 steps flagged

No circularity: decomposition is explicit modeling premise, optimum follows by standard math

full rationale

The paper introduces the Coverage-Density Trade-off by explicitly decomposing policy error into an estimation term (decreasing in per-condition density) and an extrapolation term (decreasing in coverage). From this additive structure it derives the existence of an interior optimum for fixed budget via ordinary minimization. This is a forward modeling step, not a reduction: the decomposition is posited as input, the interior optimum is its mathematical consequence, and neither is defined in terms of the other nor obtained by fitting then relabeling. No self-citations, ansatzes smuggled via prior work, or renamings of known results appear as load-bearing elements. Real-robot experiments supply external validation independent of the theoretical characterization.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the central analysis rests on the domain assumption that policy error decomposes cleanly into density and coverage terms; no explicit free parameters, invented entities, or additional axioms are stated.

axioms (1)
  • domain assumption Policy error can be decomposed into estimation (density) and extrapolation (coverage) terms that exhibit a Coverage-Density Trade-off with an interior optimum for fixed data budgets.
    Invoked directly in the formalization of the diversity trap and the characterization of optimal allocation.

pith-pipeline@v0.9.0 · 5500 in / 1459 out tokens · 39579 ms · 2026-05-11T02:21:19.756347+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

96 extracted references · 96 canonical work pages · 22 internal anchors

  1. [1]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    work page 2000

  2. [2]

    arXiv preprint arXiv:2510.01174 , year=

    Code2Video: A Code-centric Paradigm for Educational Video Generation , author=. arXiv preprint arXiv:2510.01174 , year=

    work page arXiv

  3. [3]

    Conference on Robot Learning , pages=

    Rt-2: Vision-language-action models transfer web knowledge to robotic control , author=. Conference on Robot Learning , pages=. 2023 , organization=

    work page 2023

  4. [4]

    IEEE Access , year=

    Vision-language-action models for robotics: A review towards real-world applications , author=. IEEE Access , year=

    work page

  5. [5]

    CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation

    Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation , author=. arXiv preprint arXiv:2411.19650 , year=

    work page Pith review arXiv

  6. [6]

    Black, Kevin and Brown, Noah and Driess, Danny and Esmail, Adnan and Equi, Michael and Finn, Chelsea and Fusai, Niccolo and Groom, Lachy and Hausman, Karol and Ichter, Brian and others , journal=

    work page

  7. [7]

    Black, Kevin and Brown, Noah and Darpinian, James and Dhabalia, Karan and Driess, Danny and Esmail, Adnan and Equi, Michael Robert and Finn, Chelsea and Fusai, Niccolo and Galliker, Manuel Y and others , booktitle=

    work page

  8. [8]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    Gr00t n1: An open foundation model for generalist humanoid robots , author=. arXiv preprint arXiv:2503.14734 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    OpenVLA: An Open-Source Vision-Language-Action Model

    Openvla: An open-source vision-language-action model , author=. arXiv preprint arXiv:2406.09246 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control

    Dexvla: Vision-language model with plug-in diffusion expert for general robot control , author=. arXiv preprint arXiv:2502.05855 , year=

    work page Pith review arXiv

  11. [11]

    WorldVLA: Towards Autoregressive Action World Model

    WorldVLA: Towards Autoregressive Action World Model , author=. arXiv preprint arXiv:2506.21539 , year=

    work page internal anchor Pith review arXiv

  12. [12]

    FAST: Efficient Action Tokenization for Vision-Language-Action Models

    Fast: Efficient action tokenization for vision-language-action models , author=. arXiv preprint arXiv:2501.09747 , year=

    work page internal anchor Pith review arXiv

  13. [13]

    Llada-vla: Vision language diffusion action models.arXiv preprint arXiv:2509.06932,

    Llada-vla: Vision language diffusion action models , author=. arXiv preprint arXiv:2509.06932 , year=

    work page arXiv

  14. [14]

    arXiv preprint arXiv:2507.04447 (2025) 3, 7, 14

    Dreamvla: a vision-language-action model dreamed with comprehensive world knowledge , author=. arXiv preprint arXiv:2507.04447 , year=

    work page arXiv

  15. [15]

    Rynnvla-002: A unified vision-language-action and world model.arXiv preprint arXiv:2511.17502, 2025

    RynnVLA-002: A Unified Vision-Language-Action and World Model , author=. arXiv preprint arXiv:2511.17502 , year=

    work page arXiv

  16. [16]

    Reinforcement fine-tuning of flow-matching policies for vision-language-action models, 2025

    Reinforcement Fine-Tuning of Flow-Matching Policies for Vision-Language-Action Models , author=. arXiv preprint arXiv:2510.09976 , year=

    work page arXiv

  17. [17]

    AsyncVLA: Asynchronous Flow Matching for Vision-Language-Action Models

    AsyncVLA: Asynchronous Flow Matching for Vision-Language-Action Models , author=. arXiv preprint arXiv:2511.14148 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    A Survey on Vision-Language-Action Models for Embodied AI

    A survey on vision-language-action models for embodied ai , author=. arXiv preprint arXiv:2405.14093 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics

    Smolvla: A vision-language-action model for affordable and efficient robotics , author=. arXiv preprint arXiv:2506.01844 , year=

    work page internal anchor Pith review arXiv

  20. [20]

    Flowvla: Visual chain of thought-based motion reasoning for vision-language-action models.arXiv preprint arXiv:2508.18269, 2025

    Flowvla: Visual chain of thought-based motion reasoning for vision-language-action models , author=. arXiv preprint arXiv:2508.18269 , year=

    work page arXiv

  21. [21]

    arXiv preprint arXiv:2512.07582 , year=

    See Once, Then Act: Vision-Language-Action Model with Task Learning from One-Shot Video Demonstrations , author=. arXiv preprint arXiv:2512.07582 , year=

    work page arXiv

  22. [22]

    Octo: An Open-Source Generalist Robot Policy

    Octo: An open-source generalist robot policy , author=. arXiv preprint arXiv:2405.12213 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Onetwovla: A unified vision-language-action model with adaptive reasoning

    OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning , author=. arXiv preprint arXiv:2505.11917 , year=

    work page arXiv

  24. [24]

    MolmoAct: Action Reasoning Models that can Reason in Space

    Molmoact: Action reasoning models that can reason in space , author=. arXiv preprint arXiv:2508.07917 , year=

    work page internal anchor Pith review arXiv

  25. [25]

    arXiv preprint arXiv:2512.02902 , year=

    VLA Models Are More Generalizable Than You Think: Revisiting Physical and Spatial Modeling , author=. arXiv preprint arXiv:2512.02902 , year=

    work page arXiv

  26. [26]

    Mos-vla: A vision-language-action model with one-shot skill adaptation.arXiv preprint arXiv:2510.16617, 2025

    MoS-VLA: A Vision-Language-Action Model with One-Shot Skill Adaptation , author=. arXiv preprint arXiv:2510.16617 , year=

    work page arXiv

  27. [27]

    Policy agnostic rl: Offline rl and online rl fine-tuning of any class and backbone,

    Policy agnostic rl: Offline rl and online rl fine-tuning of any class and backbone , author=. arXiv preprint arXiv:2412.06685 , year=

    work page arXiv

  28. [28]

    Self-improving vision-language-action models with data generation via residual rl.arXiv preprint arXiv:2511.00091, 2025

    Self-improving vision-language-action models with data generation via residual rl , author=. arXiv preprint arXiv:2511.00091 , year=

    work page arXiv

  29. [29]

    Conrft: A reinforcedfine-tuningmethodforvlamodelsviaconsistencypolicy.arXivpreprintarXiv:2502.05450, 2025c

    Conrft: A reinforced fine-tuning method for vla models via consistency policy , author=. arXiv preprint arXiv:2502.05450 , year=

    work page arXiv

  30. [30]

    K., and Panov, A

    Don't Blind Your VLA: Aligning Visual Representations for OOD Generalization , author=. arXiv preprint arXiv:2510.25616 , year=

    work page arXiv

  31. [31]

    Advances in Neural Information Processing Systems , volume=

    Libero: Benchmarking knowledge transfer for lifelong robot learning , author=. Advances in Neural Information Processing Systems , volume=

    work page

  32. [32]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Towards generalizable vision-language robotic manipulation: A benchmark and llm-guided 3d policy , author=. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2025 , organization=

    work page 2025

  33. [33]

    arXiv preprint arXiv:2508.10259 , year=

    Leveraging os-level primitives for robotic action management , author=. arXiv preprint arXiv:2508.10259 , year=

    work page arXiv

  34. [34]

    Nature Machine Intelligence , pages=

    Preserving and combining knowledge in robotic lifelong reinforcement learning , author=. Nature Machine Intelligence , pages=. 2025 , publisher=

    work page 2025

  35. [35]

    LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models

    Libero-plus: In-depth robustness analysis of vision-language-action models , author=. arXiv preprint arXiv:2510.13626 , year=

    work page internal anchor Pith review arXiv

  36. [36]

    Continually Evolving Skill Knowledge in Vision Language Action Model

    Continually Evolving Skill Knowledge in Vision Language Action Model , author=. arXiv preprint arXiv:2511.18085 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2506.17561 , year=

    VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models , author=. arXiv preprint arXiv:2506.17561 , year=

    work page arXiv

  38. [38]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Construct-vl: Data-free continual structured vl concepts learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  39. [39]

    Evolve-vla: Test-time training from environment feedback for vision-language-action models.arXiv preprint arXiv:2512.14666, 2025

    EVOLVE-VLA: Test-Time Training from Environment Feedback for Vision-Language-Action Models , author=. arXiv preprint arXiv:2512.14666 , year=

    work page arXiv

  40. [40]

    Leave no observation behind: Real-time correction for vla action chunks.ArXiv, abs/2509.23224, 2025

    Leave no observation behind: Real-time correction for vla action chunks , author=. arXiv preprint arXiv:2509.23224 , year=

    work page arXiv

  41. [41]

    IEEE Transactions on Cognitive and Developmental Systems , volume=

    Continual robot learning using self-supervised task inference , author=. IEEE Transactions on Cognitive and Developmental Systems , volume=. 2023 , publisher=

    work page 2023

  42. [42]

    SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning

    SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning , author=. arXiv preprint arXiv:2503.03480 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  43. [43]

    Flow Matching for Generative Modeling

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  44. [44]

    Clare: Continual learning for vision-language-action models via autonomous adapter routing and expansion.arXiv preprint arXiv:2601.09512, 2026

    CLARE: Continual Learning for Vision-Language-Action Models via Autonomous Adapter Routing and Expansion , author=. arXiv preprint arXiv:2601.09512 , year=

    work page arXiv

  45. [45]

    2022 , publisher=

    Partial differential equations , author=. 2022 , publisher=

    work page 2022

  46. [46]

    Neural computation , volume=

    Neural networks and the bias/variance dilemma , author=. Neural computation , volume=. 1992 , publisher=

    work page 1992

  47. [47]

    International journal of neural systems , volume=

    Gaussian processes for machine learning , author=. International journal of neural systems , volume=. 2004 , publisher=

    work page 2004

  48. [48]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. ICLR , volume=

    work page

  49. [49]

    IEEE Robotics and Automation Letters , volume=

    Bottom-up skill discovery from unsegmented demonstrations for long-horizon robot manipulation , author=. IEEE Robotics and Automation Letters , volume=. 2022 , publisher=

    work page 2022

  50. [50]

    Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning,

    Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning , author=. arXiv preprint arXiv:2407.01531 , year=

    work page arXiv

  51. [51]

    Parallels between vla model post-training and human motor learning: Progress, challenges, and trends.arXiv preprint arXiv:2506.20966, 2025

    Parallels Between VLA Model Post-Training and Human Motor Learning: Progress, Challenges, and Trends , author=. arXiv preprint arXiv:2506.20966 , year=

    work page arXiv

  52. [52]

    Interactive post-training for vision-language- action models, 2025

    Interactive Post-Training for Vision-Language-Action Models , author=. arXiv preprint arXiv:2505.17016 , year=

    work page arXiv

  53. [53]

    arXiv preprint arXiv:2509.19012 (2025)

    Pure vision language action (vla) models: A comprehensive survey , author=. arXiv preprint arXiv:2509.19012 , year=

    work page arXiv

  54. [54]

    Advances in Neural Information Processing Systems , volume=

    Embodiedgpt: Vision-language pre-training via embodied chain of thought , author=. Advances in Neural Information Processing Systems , volume=

    work page

  55. [55]

    MimicDreamer: Aligning human and robot demonstrations for scalable vla training

    Mimicdreamer: Aligning human and robot demonstrations for scalable vla training , author=. arXiv preprint arXiv:2509.22199 , year=

    work page arXiv

  56. [56]

    UniVLA: Learning to Act Anywhere with Task-centric Latent Actions

    Univla: Learning to act anywhere with task-centric latent actions , author=. arXiv preprint arXiv:2505.06111 , year=

    work page internal anchor Pith review arXiv

  57. [57]

    Everydayvla: A vision-language-action model for affordable robotic manipulation.arXiv preprint arXiv:2511.05397, 2025

    EveryDayVLA: A Vision-Language-Action Model for Affordable Robotic Manipulation , author=. arXiv preprint arXiv:2511.05397 , year=

    work page arXiv

  58. [58]

    Scalable vision-language-action model pretraining for robotic manipulation with real-life human activity videos.arXiv preprint arXiv:2510.21571, 2025

    Scalable vision-language-action model pretraining for robotic manipulation with real-life human activity videos , author=. arXiv preprint arXiv:2510.21571 , year=

    work page arXiv

  59. [59]

    Dita: Scaling diffusion transformer for generalist vision-language-action policy.arXiv preprint arXiv:2503.19757, 2025

    Dita: Scaling diffusion transformer for generalist vision-language-action policy , author=. arXiv preprint arXiv:2503.19757 , year=

    work page arXiv

  60. [60]

    What makes good data for alignment? a comprehensive study of automatic data selection in instruction tuning

    What makes good data for alignment? a comprehensive study of automatic data selection in instruction tuning , author=. arXiv preprint arXiv:2312.15685 , year=

    work page arXiv

  61. [61]

    LLaVA-OneVision: Easy Visual Task Transfer

    Llava-onevision: Easy visual task transfer , author=. arXiv preprint arXiv:2408.03326 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  62. [62]

    2005 , publisher=

    The organization of behavior: A neuropsychological theory , author=. 2005 , publisher=

    work page 2005

  63. [63]

    Biological cybernetics , volume=

    Mathematical formulations of Hebbian learning , author=. Biological cybernetics , volume=. 2002 , publisher=

    work page 2002

  64. [64]

    Do As I Can, Not As I Say: Grounding Language in Robotic Affordances

    Do as i can, not as i say: Grounding language in robotic affordances , author=. arXiv preprint arXiv:2204.01691 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  65. [65]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0 , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    work page 2024

  66. [66]

    Evaluating Real-World Robot Manipulation Policies in Simulation

    Evaluating real-world robot manipulation policies in simulation , author=. arXiv preprint arXiv:2405.05941 , year=

    work page internal anchor Pith review arXiv

  67. [67]

    Gemini Robotics: Bringing AI into the Physical World

    Gemini robotics: Bringing ai into the physical world , author=. arXiv preprint arXiv:2503.20020 , year=

    work page internal anchor Pith review arXiv

  68. [68]

    VLA-Arena: An open-source framework for benchmarking vision-language-action models.arXiv preprint arXiv:2512.22539, 2025

    VLA-Arena: An Open-Source Framework for Benchmarking Vision-Language-Action Models , author=. arXiv preprint arXiv:2512.22539 , year=

    work page arXiv

  69. [69]

    arXiv preprint arXiv:2508.02062 , year=

    Ricl: Adding in-context adaptability to pre-trained vision-language-action models , author=. arXiv preprint arXiv:2508.02062 , year=

    work page arXiv

  70. [70]

    arXiv preprint arXiv:2512.01715 , year=

    DiG-Flow: Discrepancy-Guided Flow Matching for Robust VLA Models , author=. arXiv preprint arXiv:2512.01715 , year=

    work page arXiv

  71. [71]

    arXiv preprint arXiv:2509.19752 , year=

    Beyond human demonstrations: Diffusion-based reinforcement learning to generate data for vla training , author=. arXiv preprint arXiv:2509.19752 , year=

    work page arXiv

  72. [72]

    arXiv preprint arXiv:2512.00903 (2025)

    SwiftVLA: Unlocking Spatiotemporal Dynamics for Lightweight VLA Models at Minimal Overhead , author=. arXiv preprint arXiv:2512.00903 , year=

    work page arXiv

  73. [73]

    IEEE Robotics and Automation Letters , year=

    Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation , author=. IEEE Robotics and Automation Letters , year=

    work page

  74. [74]

    Correctnav: Self-correction flywheel empowers vision-language-action navigation model.arXiv preprint arXiv:2508.10416, 2025

    CorrectNav: Self-Correction Flywheel Empowers Vision-Language-Action Navigation Model , author=. arXiv preprint arXiv:2508.10416 , year=

    work page arXiv

  75. [75]

    Wmpo: World model-based policy optimization for vision-language-action models.arXiv preprint arXiv:2511.09515, 2025

    Wmpo: World model-based policy optimization for vision-language-action models , author=. arXiv preprint arXiv:2511.09515 , year=

    work page arXiv

  76. [76]

    RESample: A Robust Data Augmentation Framework via Exploratory Sampling for Robotic Manipulation

    Resample: A robust data augmentation framework via exploratory sampling for robotic manipulation , author=. arXiv preprint arXiv:2510.17640 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  77. [77]

    arXiv preprint arXiv:2310.01362 , year=

    Robot fleet learning via policy merging , author=. arXiv preprint arXiv:2310.01362 , year=

    work page arXiv

  78. [78]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Lotus: Continual imitation learning for robot manipulation through unsupervised skill discovery , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    work page 2024

  79. [79]

    arXiv preprint arXiv:2506.09623 , year=

    Analytic Task Scheduler: Recursive Least Squares Based Method for Continual Learning in Embodied Foundation Models , author=. arXiv preprint arXiv:2506.09623 , year=

    work page arXiv

  80. [80]

    arXiv preprint arXiv:2505.15424 , year=

    Gated Integration of Low-Rank Adaptation for Continual Learning of Language Models , author=. arXiv preprint arXiv:2505.15424 , year=

    work page arXiv

Showing first 80 references.