Pith. sign in

REVIEW 3 major objections 4 minor 4 cited by

A vision-language-action model can learn tactile-aware grasping from a distilled teacher and run without any touch sensor at deployment.

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 · deepseek-v4-flash

2026-08-04 05:46 UTC pith:YYXHYUKG

load-bearing objection Real system, plausible idea, but the headline claim overreaches: the abstract's token-prediction mechanism isn't in the method, and 20-trial results don't support 'consistently outperforming.' the 3 major comments →

arxiv 2603.15257 v2 pith:YYXHYUKG submitted 2026-03-16 cs.RO

HapticVLA: Contact-Rich Manipulation via Vision-Language-Action Model without Inference-Time Tactile Sensing

Konstantin Gubernatorov , Mikhail Sannikov , Ilya Mikhalchuk , Egor Kuznetsov , Makar Artemov , Ogunwoye Faith Oluwatobi , Marcelino Fernando , Artem Asanov
show 2 more authors
Ziang Guo Dzmitry Tsetserukou
This is my paper
classification cs.RO
keywords vision-language-action modeltactile distillationcontact-rich manipulationflow matchingsafety-aware rewardknowledge distillationsensor-free deploymentfragile object manipulation
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.

The paper tries to show that tactile awareness is a trainable, distillable capability, not a hardware requirement. It trains a tactile-conditioned teacher policy with flow matching weighted by a safety reward that penalizes excessive grasping force, then distills the teacher's action chunks into a student VLA that receives only vision, language, and joint-state inputs. The student achieves a mean success rate of 86.7% across three fragile-object pick-and-place tasks, beating baselines that receive actual tactile feedback at inference. If true, contact-rich manipulation can be deployed on cheaper, sensor-free robots without sacrificing haptic-safe behavior.

Core claim

HapticVLA claims that tactile-aware behavior can be transferred into a VLA that never touches a sensor at runtime. A tactile-conditioned flow-matching teacher is trained with reward weights derived from offline safety penalties (over-force, slip, pressure concentration, pad asymmetry). Its predicted action chunks, blended 50/50 with ground-truth demonstrations, train a conventional VLA student that receives only vision, language, and joint state. The student inherits the teacher's proprioception projection but drops the tactile encoder. On three fragile-object tasks (jar, waffles, egg) it reaches 86.7% mean success, beating baselines that receive live tactile feedback.

What carries the argument

The load-bearing pair is (SA-RWFM, TD). SA-RWFM weights each flow-matching training sample by an exponentiated, clipped advantage score built from per-step tactile penalties and episode risk, steering the generative policy toward low-force, safe trajectories. TD then copies all teacher weights except the tactile encoder, initializes the student's state projection from the teacher's proprioception-only columns, and trains the student with a flow-matching loss on blended action targets (half ground truth, half teacher prediction). This lets a sensor-free student reproduce the teacher's force-modulated action distribution.

Load-bearing premise

The central claim depends on the premise that a student trained only to imitate the teacher's action chunks inherits the teacher's tactile awareness, rather than merely learning a vision-to-action mapping that happens to succeed on the tested objects.

What would settle it

Run the distilled student on objects that are visually identical but differ in fragility — e.g., a real egg versus a ceramic egg of the same color and texture. If success drops to chance, the student has no true tactile awareness. A direct check would be to compare the student's internal representations against the teacher's tactile embeddings on held-out episodes; if they share no mutual information with force measurements, there is no distilled tactile token.

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

If this is right

  • Contact-rich policies can run on robot arms with no fingertip tactile sensors, cutting hardware cost and compatibility constraints.
  • Offline safety-reward weighting can outperform injecting raw tactile signals at inference, according to the reported results.
  • Distillation runs fully offline from one teacher pass, so no teacher is needed during training or deployment of the student.
  • The student requires zero architectural changes, so the same recipe transfers to other VLA backbones.

Where Pith is reading between the lines

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

  • If the claim holds, tactile information is largely redundant with vision and state in these tasks; a testable extension is to curate contact-rich datasets by annotating force limits offline rather than instrumenting every robot.
  • A natural probe: train a linear classifier on the student's internal activations to see whether it encodes force estimates, or whether the gains come from task-specific vision-to-action shortcuts.
  • The paper's abstract promises a distilled tactile token, but the method actually blends teacher action chunks with ground truth; if action-level imitation is sufficient, simpler distillation objectives could transfer other sensor modalities (e.g., audio or proprioception) the same way.

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 / 4 minor

Summary. The paper introduces HapticVLA, a two-stage method for contact-rich VLA manipulation without tactile sensors at deployment. In the first stage (SA-RWFM), a flow-matching action expert is fine-tuned on an offline dataset using per-step and episode-level tactile safety rewards derived from force maps; in the second stage (TD), the teacher's action predictions are precomputed and the student is trained on a blend of ground-truth and teacher action targets. The paper reports real-world experiments on three pick-and-place tasks (jar, waffles, egg), with 20 trials per model/task, claiming an 86.7% mean success rate over several VLA baselines and over tactile-feedback ablations. Code, models, data, and a digital twin are said to be released.

Significance. If the mechanism and results held, the paper would be a useful practical contribution: it would show that tactile-aware safe manipulation can be precomputed into a vision/state-only VLA, reducing hardware cost while retaining contact-rich performance. The real-world evaluation, the explicit inclusion of faulty demonstrations, and the released assets are strengths. However, the central scientific claim — that tactile-aware reasoning is distilled as a 'tactile token' — is not implemented in the method as written, and the statistical evidence is too weak to support the strong 'consistently outperforming' language. If reframed as action-level distillation, the contribution is more modest and would need additional evidence to support tactile-awareness claims.

major comments (3)
  1. [III.B, Eqs. (21)-(25)] The abstract and Fig. 2 state that TD 'distills a compact tactile token' and trains a student to predict this token from vision and state. No such token appears in the method. Stage 1 stores teacher action chunks â_i; Stage 3 blends these with ground-truth actions a_GT via Eq. (24) and regresses the flow velocity toward the blended action via Eq. (25). The student is thus trained to imitate actions, not to predict any tactile representation. The claim that the student is tactile-aware is therefore an interpretation, not a demonstrated mechanism. To support the current claim, the paper should either define and train a tactile-token prediction head, or reframe the contribution as action-level distillation and provide a direct probe — e.g., held-out force/tactile prediction accuracy or a decoupling experiment — showing that the student encodes tactile information beyond imitation.
  2. [Table I / IV.C] The abstract's 'consistently outperforming' claim is contradicted by Table I: HapticVLA w/ TD scores 15/20 on Jar, while the tactile-feedback baseline w/o TD async scores 16/20; moreover, w/ TD async has a lower mean than w/o TD async (80.0 vs 81.7). With only 20 trials per condition, a one-trial difference is within binomial sampling noise. No error bars, confidence intervals, or significance tests are reported. The claim should either be limited to mean success or supported by statistical analysis.
  3. [III.A, Eqs. (7)-(10)] SA-RWFM depends on many hand-set coefficients and thresholds: λ_high, λ_low, λ_peak, λ_conc, λ_asym, λ_slip; R_step_scale, R_succ, R_drop, R_damage, R_risk; and the calibrated bounds f_min/f_max, p_max, c_max, δ. The paper gives no numerical values for most of these and no sensitivity analysis. Since the core claim is that tactile reward weighting improves the teacher and that this improvement transfers through TD, the lack of robustness evidence makes it difficult to attribute the reported 86.7% mean success to the specific reward design rather than to generic reward-weighted flow matching. At minimum, report the values and one sensitivity study over the principal penalty weights.
minor comments (4)
  1. [General] Typos and wording issues: 'fine-tine' in Section I, 'infernece' in Section IV.D, 'Waflles' in Fig. 6, and 'the the' in Section IV.C. 'criteria' should be 'criterion'. A careful proofread is needed.
  2. [Fig. 1 / Fig. 2 captions] Fig. 1 states that the model 'processes tactile reward' alongside vision/language/state, but at inference HapticVLA does not receive tactile reward. The caption should clarify that tactile reward is used only in offline teacher training, not in the deployed student.
  3. [III.A.1 / III.A.3] Several hyperparameters introduced in Eqs. (7)-(10) are never given numerical values. The text lists only α, w_min, w_max, β, c_A, and γ. Since the project website is promised to include code, the paper itself should still provide a table of all reward coefficients and calibration thresholds to make the method reproducible from the text.
  4. [IV.C] The claim that X-VLA and VLA-0 achieve zero-percent success is striking, but Fig. 6 appears to show only bars without exact values or trial counts. Reporting the raw counts in the text or a table would strengthen the comparison.

Circularity Check

0 steps flagged

No derivation-level circularity; the tactile-token mechanism in the abstract is not implemented by Eqs. (24)-(25), but that is a correctness/validity gap, not a circular reduction.

full rationale

The claimed derivation chain is: SA-RWFM reweights the flow-matching loss with tactile-derived rewards (Eqs. 7-20); TD generates teacher action chunks (Eq. 21), blends them with ground-truth actions (Eq. 24), and trains the student by regressing the flow velocity toward the blended action target (Eq. 25). None of these steps fits a parameter to the evaluation metric and then reports that metric as a prediction. The central evaluation is external: 20 real-world trials per task, with success defined by object integrity and placement, independent of the fitted reward or teacher targets. No self-citation is load-bearing; FD-VLA [12] is cited as external related work, not as the authors' own prior theorem. The strongest concern is that the abstract says the student is trained to predict a compact tactile token from vision and state, but no such token is defined, predicted, or aligned in the method; Eq. 25 only regresses to blended action chunks. That is a mechanism/construct-validity gap and an unsupported inference from task success to tactile awareness, but it is not an equivalence-by-construction and does not make the empirical result circular. Under the given circularity rubric, the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 1 invented entities

The central claim rests on several hand-set reward and training parameters that are not reported or ablated, on the assumption that tactile map statistics capture safety, and on the unverified transfer of digital-twin and teacher-generated targets. The promised tactile token is an invented entity that never appears in the implemented method.

free parameters (7)
  • tactile reward penalty weights (λ_high, λ_low, λ_peak, λ_conc, λ_asym, λ_slip) = not reported
    Eq. (7) defines the per-step reward; values are not given and no ablation shows sensitivity. They determine the teacher's notion of safety and thus what is distilled.
  • safe force band [f_min, f_max], peak-pressure limit p_max, concentration limit c_max, asymmetry tolerance δ = not reported (dataset quantiles)
    Calibrated from contact statistics in Section III.A.1.a; these thresholds shape rewards and are not independently validated.
  • dataset-level tactile normalization scale (99th percentile) = not reported
    Used to normalize raw tactile maps before reward calculation (Section III.A.1.a).
  • episode reward coefficients (R_step_scale, R_succ, R_drop, R_damage, R_risk) = not reported
    Eq. (10) combines step reward, success, drop, damage, and risk; values are unspecified but affect the teacher's reward weighting.
  • SA-RWFM weighting hyperparameters (α=0.25, w_min=0.25, w_max=4.0, β=0.7, c_A=6, γ=0.99) = stated in text
    Chosen by hand in Section III.A.4; no sensitivity analysis is reported. These determine how strongly reward reweights the imitation loss.
  • α warm-up schedule and λ_anchor schedule = not reported
    Section III.A.5 says a warm-up schedule gradually increases α and adjusts λ_anchor but gives no details; these control the stability of SA-RWFM fine-tuning.
  • blending coefficient α in Tactile Distillation = 0.5
    Eq. (24) blends ground-truth and teacher actions; α=0.5 is set by hand with no ablation on α.
axioms (5)
  • domain assumption Flow-matching action prediction and RWFM reweighting from prior work (Pfrommer et al. [15]) transfer to SmolVLA fine-tuning without modification.
    Section III.A.4 builds on prior RWFM; assumes the reward-weighted objective improves real-world success rather than causing mode collapse.
  • domain assumption Tactile map statistics (mean force, peak pressure, concentration, CoP) are a faithful proxy for grasp safety and object damage.
    Section III.A.1 uses these proxies to compute rewards; no calibration against actual damage beyond binary episode outcomes.
  • domain assumption The pretrained SmolVLA and the 310 real + 1000 simulated episodes are sufficient to fine-tune baselines and HapticVLA fairly.
    Section IV.C fine-tunes all models on the same in-house dataset; zero success for X-VLA/VLA-0 suggests the comparison may be fragile or unfair.
  • domain assumption Digital twin episodes transfer to the real system.
    Section III.C adds 1000 Isaac Sim episodes; the paper does not ablate sim vs real or validate sim-to-real transfer.
  • domain assumption Teacher action targets generated offline are high-quality soft labels for distillation.
    Section III.B.2 generates targets by teacher inference; if teacher errors correlate with fragile-object failures, blending α=0.5 may bake those errors into the student.
invented entities (1)
  • distilled tactile token no independent evidence
    purpose: A compact tactile representation the student is supposed to predict from vision/state to preserve haptic reasoning (Abstract, Fig. 2).
    No such token is defined or predicted in Section III.B; the student is trained to imitate action chunks, and no metric verifies that a tactile latent is recovered.

pith-pipeline@v1.3.0-alltime-deepseek · 11071 in / 13239 out tokens · 125440 ms · 2026-08-04T05:46:24.042102+00:00 · methodology

0 comments
read the original abstract

Tactile sensing is a crucial capability for Vision-Language-Action (VLA) architectures, as it enables dexterous and safe manipulation in contact-rich tasks. However, reliance on dedicated tactile hardware increases cost and reduces reproducibility across robotic platforms. We argue that tactile-aware manipulation can be learned offline and deployed without direct haptic feedback at inference. To this end, we present HapticVLA, which proceeds in two tightly coupled stages: Safety-Aware Reward-Weighted Flow Matching (SA-RWFM) and Tactile Distillation (TD). SA-RWFM trains a flow-matching action expert that incorporates precomputed, safety-aware tactile rewards penalizing excessive grasping force and suboptimal grasping trajectories. TD further transfers this tactile-aware capability into a conventional VLA: we distill a compact tactile token from the SA-RWFM teacher and train a student VLA to predict that token from vision and state modalities, enabling tactile-aware action generation at inference without requiring on-board tactile sensors. This design preserves contact-rich tactile-aware reasoning within VLA while removing the need for on-board tactile sensors during deployment. On real-world experiments, HapticVLA achieves a mean success rate of 86.7%, consistently outperforming baseline VLAs - including versions provided with direct tactile feedback during inference.

Figures

Figures reproduced from arXiv: 2603.15257 by Artem Asanov, Dzmitry Tsetserukou, Egor Kuznetsov, Ilya Mikhalchuk, Konstantin Gubernatorov, Makar Artemov, Marcelino Fernando, Mikhail Sannikov, Ogunwoye Faith Oluwatobi, Ziang Guo.

Figure 1
Figure 1. Figure 1: HapticVLA is the first VLA model to enable tactile-aware [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Framework overview of HapticVLA. After dataset collection we perform offline tactile reward calculation for each episode based on manipulator [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Our dataset collection setups: (Left) Real robotics platform; and (Right) Digital twin in Isaac Sim [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (Left) Tactile array with 100 taxels. (Right) Example of 10 x 10 tactile map from the array during manipulation. Red/orange regions indicate high contact forces and blue regions low contact forces. To further increase the scale of the dataset and improve the robustness to pose variations, as shown in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: For our robotics platform we use two SO-101 robot arms as the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Success rates for three contact-rich manipulation tasks: Jar, Waflles, and Egg pick-and-place. Results are averaged over 20 evaluations per task. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Human-Centric Transferable Tactile Pre-Training for Dexterous Robotic Manipulation

    cs.RO 2026-07 unverdicted novelty 6.0

    Introduces H-Tac human tactile-action dataset and TTP pre-training that unifies spaces and predicts future tactile signals to improve robotic dexterous manipulation transfer.

  2. TORL-VLA: Tactile Guided Online Reinforcement Learning for Contact-Rich Manipulation

    cs.RO 2026-06 unverdicted novelty 6.0

    TORL-VLA couples a tactile wrench-aware VLA policy with a lightweight online RL module and an intervention-censored critic to improve success and efficiency on contact-rich robotic tasks.

  3. TacCoRL: Integrating Tactile Feedback into VLA via Simulation

    cs.RO 2026-06 unverdicted novelty 5.0

    TacCoRL integrates tactile feedback into VLA policies via real-aligned simulation co-training and RL, raising average success from 50% to 72.5% on four bimanual contact-rich tasks with direct real-robot transfer.

  4. AetheRock: An Arm-Worn Robot Teaching System for Force-Guided Vision-Tactile Learning

    cs.RO 2026-06 unverdicted novelty 5.0

    Presents arm-worn AetheRock hardware for multi-modal data collection and ForceVT learning method to improve tactile inference robustness despite sensor variations.

Reference graph

Works this paper leans on

35 extracted references · 11 linked inside Pith · cited by 4 Pith papers

  1. [1]

    Open- vla: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi,et al., “Open- vla: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    π0: 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. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, J. Tanner, Q. Vuong, A. Walling, H. Wang, and U. Zhilinsky, “π0: A vision-language-action flow model for general robot control,”

  3. [3]

    Smolvla: A vision-language-action model for affordable and efficient robotics,

    M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafioti,et al., “Smolvla: A vision-language-action model for affordable and efficient robotics,”arXiv preprint arXiv:2506.01844, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Vtla: Vision- tactile-language-action model with preference learning for insertion manipulation,

    C. Zhang, P. Hao, X. Cao, X. Hao, S. Cui, and S. Wang, “Vtla: Vision- tactile-language-action model with preference learning for insertion manipulation,”arXiv preprint arXiv:2505.09577, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Mla: A multisensory language-action model for multimodal understanding and forecasting in robotic manipulation,

    Z. Liu, J. Liu, J. Xu, N. Han, C. Gu, H. Chen, K. Zhou, R. Zhang, K. C. Hsieh, K. Wu,et al., “Mla: A multisensory language-action model for multimodal understanding and forecasting in robotic manipulation,” arXiv preprint arXiv:2509.26642, 2025

    arXiv 2025

  6. [6]

    Omnivtla: Vision-tactile-language-action model with semantic-aligned tactile sensing,

    Z. Cheng, Y . Zhang, W. Zhang, H. Li, K. Wang, L. Song, and H. Zhang, “Omnivtla: Vision-tactile-language-action model with semantic-aligned tactile sensing,”arXiv preprint arXiv:2508.08706, 2025

    arXiv 2025

  7. [7]

    Vla-touch: Enhanc- ing vision-language-action models with dual-level tactile feedback,

    J. Bi, K. Y . Ma, C. Hao, M. Z. Shou, and H. Soh, “Vla-touch: Enhanc- ing vision-language-action models with dual-level tactile feedback,” arXiv preprint arXiv:2507.17294, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Bitla: A bimanual tactile-language-action model for contact-rich robotic manipulation,

    S. Yang, H. Li, J. Hu, S. Zhang, G. Yao, Z. Ni, and B. Fang, “Bitla: A bimanual tactile-language-action model for contact-rich robotic manipulation,” inProceedings of the 1st International Workshop on Multi-Sensorial Media and Applications, ser. MSMA ’25. New Fig. 6. Success rates for three contact-rich manipulation tasks: Jar, Waflles, and Egg pick-and-p...

    arXiv 2025

  9. [9]

    Beyond sight: Finetuning generalist robot policies with heterogeneous sensors via language grounding,

    J. Jones, O. Mees, C. Sferrazza, K. Stachowicz, P. Abbeel, and S. Levine, “Beyond sight: Finetuning generalist robot policies with heterogeneous sensors via language grounding,” in2025 IEEE Inter- national Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 5961–5968

    2025

  10. [10]

    Forcevla: Enhancing vla models with a force-aware moe for contact-rich manipulation,

    J. Yu, H. Liu, Q. Yu, J. Ren, C. Hao, H. Ding, G. Huang, G. Huang, Y . Song, P. Cai,et al., “Forcevla: Enhancing vla models with a force-aware moe for contact-rich manipulation,”arXiv preprint arXiv:2505.22159, 2025

    arXiv 2025

  11. [11]

    Tactile-force alignment in vision-language-action models for force- aware manipulation,

    Y . Huang, P. Lin, W. Li, D. Li, J. Li, J. Jiang, C. Xiao, and Z. Jiao, “Tactile-force alignment in vision-language-action models for force- aware manipulation,”arXiv preprint arXiv:2601.20321, 2026

    arXiv 2026

  12. [12]

    Fd-vla: Force-distilled vision-language-action model for contact-rich manipulation,

    R. Zhao, W. Wang, Y . Ma, X. Li, F. E. Tay, M. H. Ang Jr, and H. Zhu, “Fd-vla: Force-distilled vision-language-action model for contact-rich manipulation,”arXiv preprint arXiv:2602.02142, 2026

    arXiv 2026

  13. [13]

    Tactile- vla: unlocking vision-language-action model’s physical knowledge for tactile generalization,

    J. Huang, S. Wang, F. Lin, Y . Hu, C. Wen, and Y . Gao, “Tactile- vla: unlocking vision-language-action model’s physical knowledge for tactile generalization,”arXiv preprint arXiv:2507.09160, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Low-fidelity visuo-tactile pre-training improves vision-only manipulation performance,

    S. Gano, A. George, and A. B. Farimani, “Low-fidelity visuo-tactile pre-training improves vision-only manipulation performance,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 15 983–15 989

    2025

  15. [15]

    Reinforcement learning for flow-matching policies,

    S. Pfrommer, Y . Huang, and S. Sojoudi, “Reinforcement learning for flow-matching policies,”arXiv preprint arXiv:2507.15073, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Online reward-weighted fine-tuning of flow matching with wasserstein reg- ularization,

    J. Fan, S. Shen, C. Cheng, Y . Chen, C. Liang, and G. Liu, “Online reward-weighted fine-tuning of flow matching with wasserstein reg- ularization,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  17. [17]

    Flowrl: Matching reward distributions for llm reasoning,

    X. Zhu, D. Cheng, D. Zhang, H. Li, K. Zhang, C. Jiang, Y . Sun, E. Hua, Y . Zuo, X. Lv,et al., “Flowrl: Matching reward distributions for llm reasoning,”arXiv preprint arXiv:2509.15207, 2025

    arXiv 2025

  18. [18]

    Balancing signal and variance: Adaptive offline rl post-training for vla flow models,

    H. Zhang, S. Zhang, J. Jin, Q. Zeng, Y . Qiao, H. Lu, and D. Wang, “Balancing signal and variance: Adaptive offline rl post-training for vla flow models,”arXiv preprint arXiv:2509.04063, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Flowram: Grounding flow matching policy with region-aware mamba framework for robotic manipulation,

    S. Wang, L. Wang, S. Zhou, J. Tian, J. Li, H. Sun, and W. Tang, “Flowram: Grounding flow matching policy with region-aware mamba framework for robotic manipulation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 12 176–12 186

    2025

  20. [20]

    Reinflow: Fine-tuning flow matching policy with online reinforcement learning,

    T. Zhang, C. Yu, S. Su, and Y . Wang, “Reinflow: Fine-tuning flow matching policy with online reinforcement learning,”arXiv preprint arXiv:2505.22094, 2025

    arXiv 2025

  21. [21]

    Meta-world: A benchmark and evaluation for multi-task and meta-reinforcement learning,

    T. Yu, D. Quillen, Z. He, R. Julian, K. Hausman, C. Finn, and S. Levine, “Meta-world: A benchmark and evaluation for multi-task and meta-reinforcement learning,” inConference on Robot Learning (CoRL), 2019

    2019

  22. [22]

    Learning complex dexterous manipulation with deep reinforcement learning and demonstrations,

    A. Rajeswaran, V . Kumar, A. Gupta, and S. Levine, “Learning complex dexterous manipulation with deep reinforcement learning and demonstrations,” inRobotics: Science and Systems (RSS), 2018

    2018

  23. [23]

    X- distill: Cross-architecture vision distillation for visuomotor learning,

    M. Shao, F. Zhang, G. Zhang, B. Cheng, Z. Xue, and H. Xu, “X- distill: Cross-architecture vision distillation for visuomotor learning,” arXiv preprint arXiv:2601.11269, 2026

    arXiv 2026

  24. [24]

    Libero: Benchmarking knowledge transfer in lifelong robot learning,

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “Libero: Benchmarking knowledge transfer in lifelong robot learning,” inNeurIPS, 2023

    2023

  25. [25]

    Simpler: Mapping the gap between vla sim- ulation and real-world evaluation,

    X. Li, K. Zhang,et al., “Simpler: Mapping the gap between vla sim- ulation and real-world evaluation,”arXiv preprint arXiv:2405.05941, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks,

    O. Mees, L. Hermann, E. Rosete-Beas, and W. Burgard, “Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks,”IEEE Robotics and Automation Letters (RA- L), 2022

    2022

  27. [27]

    Actdistill: General action-guided self-derived distillation for efficient vision-language- action models,

    W. Ye, T. Wang, L. Zhu, F. Li, and G. Yang, “Actdistill: General action-guided self-derived distillation for efficient vision-language- action models,”arXiv preprint arXiv:2511.18082, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    Dysl-vla: Effi- cient vision-language-action model inference via dynamic-static layer- skipping for robot manipulation,

    Z. Yang, Y . p. Qi, T. Xie, B. Yu, S. Liu, and M. Li, “Dysl-vla: Effi- cient vision-language-action model inference via dynamic-static layer- skipping for robot manipulation,”arXiv preprint arXiv:2602.22896, 2026

    arXiv 2026

  29. [29]

    Robobrain: A unified brain model for robotic manipulation from abstract to concrete,

    Y . Ji, H. Tan, J. Shi, X. Hao, Y . Zhang, H. Zhang, P. Wang, M. Zhao, Y . Mu, P. An,et al., “Robobrain: A unified brain model for robotic manipulation from abstract to concrete,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  30. [30]

    Move-kd: Knowledge distillation for vlms with mixture of visual encoders,

    J. Cao, Y . Zhang, T. Huang, M. Lu, Q. Zhang, R. An, N. Ma, and S. Zhang, “Move-kd: Knowledge distillation for vlms with mixture of visual encoders,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  31. [31]

    Fedvla: Federated vision-language-action learning with dual gating mixture-of-experts for robotic manipulation,

    C. Miao, T. Chang, M. Wu, H. Xu, C. Li, L. Ming, and X. Wang, “Fedvla: Federated vision-language-action learning with dual gating mixture-of-experts for robotic manipulation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025

    2025

  32. [32]

    Lerobot: State-of-the-art machine learning for real-world robotics in pytorch,

    R. Cadene, S. Alibert, A. Soare, Q. Gallouedec, A. Zouitine, S. Palma, P. Kooijmans, M. Aractingi, M. Shukor, D. Aubakirova, M. Russi, F. Capuano, C. Pascal, J. Choghari, J. Moss, and T. Wolf, “Lerobot: State-of-the-art machine learning for real-world robotics in pytorch,” https://github.com/huggingface/lerobot, 2024

    2024

  33. [33]

    X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model,

    J. Zheng, J. Li, Z. Wang, D. Liu, X. Kang, Y . Feng, Y . Zheng, J. Zou, Y . Chen, J. Zeng,et al., “X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model,”arXiv preprint arXiv:2510.10274, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Vla-0: Building state-of-the-art vlas with zero modification,

    A. Goyal, H. Hadfield, X. Yang, V . Blukis, and F. Ramos, “Vla-0: Building state-of-the-art vlas with zero modification,”arXiv preprint arXiv:2510.13054, 2025

    arXiv 2025

  35. [2024]

    Available: https://arxiv.org/abs/2410.24164

    [Online]. Available: https://arxiv.org/abs/2410.24164

    Pith/arXiv arXiv