Pith. sign in

REVIEW 3 major objections 4 minor 27 references

Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Two failure modes explain VLA contact misses; fix lifts success to 66%

desk verdict Good empirical paper with a solid force-injection story and careful ablations, but the precision-failure diagnosis is statistically weaker than advertised and the headline decomposition is not secure. read the letter →

arxiv 2608.01402 v1 pith:QE2SI5WF submitted 2026-08-02 cs.RO cs.AI

Carlota Parés-Morlans , Nils Kuhn , Isabel Liu , Alberta Longhini , Jeannette Bohg This is my paper
classification cs.ROcs.AI
keywords contact-richmanipulationvision-language-actionmodelsflowmatchingnoisescheduleforcesensingimitationlearningprecisionfailures
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Vision-language-action models fail in contact-rich manipulation for two distinct reasons, this paper argues: a training-schedule mismatch that starves the fine-correction regime, and a failure to exploit the temporal structure of force signals. The paper proposes FACT, which fixes the first with a Logit-Normal noise schedule and the second with time-aware force injection. Across five real-world tasks and nearly 2,500 rollouts, FACT raises average success from 39% (plain pi0.5) to 66%, beating the best prior baseline by about 25 percentage points. The two mechanisms are shown to be causally independent through ablations.

What carries the argument

The noise level $\tau$ in the flow-matching objective. The paper treats $\tau$'s sampling distribution as the first-class bottleneck: shifting it from Beta to Logit-Normal with location parameter $m=1.5$ reallocates training signal to the $\tau<0.2$ contact-correction regime. The second mechanism is time-aware force injection, which combines a shared causal temporal encoder over $H=30$ windows of force history, a current-force modulation of AdaRMSNorm scales in every transformer layer, and a gradient gate (threshold $\delta=0.5$ N) that blocks learning from near-zero force steps. Together these carry the argument that each failure mode has a targeted fix.

What would settle it

Re-running the five-task comparison with the original published ForceVLA and TA-VLA code or checkpoints under the same 40-rollout protocol: if either original baseline matches or exceeds FACT's 66% average success, the claimed margin collapses. Alternatively, a precision-critical task where success rate does not rise as the LN location parameter m moves from -1.5 to 1.5 would undermine the starvation diagnosis.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that flow-based VLAs underperform on contact-rich tasks because of two causally distinct failure modes. Precision failures stem from the Beta noise schedule used in flow matching, which allocates only 8.9% of gradient signal to the low-noise regime $\tau<0.2$ where sub-millimeter corrections are generated. Force failures stem from the structure of force signals—sparsity, temporal dynamics, and state-dependent sensitivity—which naive concatenation fails to capture. FACT replaces Beta with a Logit-Normal schedule (location $m=1.5$) that multiplies low-noise gradient signal sixfold, and injects force through a causal temporal encoder plus per-layer RMS modulat

Load-bearing premise

The headline margin depends on the authors' re-implementations of ForceVLA and TA-VLA on the pi0.5 backbone; if those re-implementations are weaker than the original published methods, the reported 25-point advantage is overstated.

Editorial extensions

If this is right

  • Replacing the Beta noise schedule with Logit-Normal improves precision-critical tasks by up to 20 percentage points and force-critical tasks by up to 45 points, without adding parameters or data.
  • Time-aware force injection adds up to 22.5 percentage points on force-critical tasks, with force history the most important component: removing it drops button push by 62.5 points.
  • The Logit-Normal schedule transfers as a drop-in fix to other flow-based VLA backbones (pi0.5 and pi0) and to force-augmented baselines, improving all of them.
  • The two mechanisms are causally independent; ablations show additive gains, and a noise-substitution test shows FACT genuinely exploits force while one baseline does not.
  • The decomposition implies that future contact-rich VLA work should treat training schedule and force structure as separate axes, not just architecture and data.
  • If the decomposition is correct, some 'force reasoning' gains in prior methods may actually come from improved low-noise training rather than from understanding force.
  • The force-history dominance suggests that contact dynamics are best captured temporally; architectures compressing force into a single token may be fundamentally limited.
  • A direct testable extension is applying the schedule-only fix to contact-rich tasks without any force sensor, to separate training effects from sensing effects.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper aims to explain why vision-language-action (VLA) models fail in contact-rich manipulation, proposing two causally distinct failure modes. Precision failures are attributed to a flow-matching training mismatch: standard Beta noise schedules under-train the low-noise denoising regime responsible for small contact corrections. Force failures are attributed to the sparsity, temporal structure, and contact-dependent salience of force signals, which naive force concatenation does not exploit. The authors propose FACT, combining a Logit-Normal (LN) noise schedule with a time-aware force injection mechanism that modulates the action expert via AdaRMSNorm and prepends force-history tokens. They evaluate on five real-robot tasks with roughly 2,500 rollouts, reporting 66% average success for FACT versus 40.5% for their re-implemented ForceVLA baseline. Ablations and a noise-substitution experiment are used to attribute gains to the two mechanisms.

Significance. If the two-failure-mode decomposition holds, it would be a useful conceptual contribution to the growing literature on force-augmented VLAs, shifting attention from purely architectural fixes to training-distribution and signal-structure issues. The proposed interventions are simple, add few parameters, and appear transferable across flow-matching backbones. The paper also provides substantial real-robot evaluation, statistical significance testing, and a noise-substitution control, which are strengths. However, the central claim of a precision-specific failure mode is not securely established by the reported evidence: the gains on the two precision-critical tasks are not individually significant, and the schedule hyperparameter is tuned on one of those tasks. The headline comparison also depends on author re-implementations of prior baselines without code release or parity checks. These issues make the current version convincing as an engineering recipe but not yet as a validated causal decomposition.

major comments (3)
  1. [Section 7, Table 1, and Appendix F] The precision-failure diagnosis is not supported at the claimed significance level. For the two precision-critical tasks, π0.5+LN vs π0.5 gives plug +20 pp (p=.055) and USB +10 pp (p=.249), both non-significant at α=.05. The only individually significant LN gains are on tasks the paper labels force-critical: button +45 pp (p<.001) and key +22.5 pp (p=.020). This does not isolate a precision-specific training mismatch. Moreover, Appendix F sweeps the LN location parameter m on plug insertion, the same task used as the headline precision result in Table 1, and selects m=1.5. The plug result is therefore in-sample tuned, not an independent confirmation. Since the two-mode decomposition is the paper's central contribution, this needs a task-agnostic schedule choice or an out-of-sample validation, or the claim must be weakened.
  2. [Section 6 and Appendix E] The headline margin over prior baselines rests on the authors' own re-implementations of ForceVLA and TA-VLA on the π0.5 backbone. No code, checkpoints, or comparison with the original implementations are provided. If the re-implementations are weaker than the published versions, the reported 'best prior baseline' of 40.5% is understated and the 25.5 pp advantage is inflated. Additionally, the 'π0 backbone' results in Table 4 use a modified π0 action expert with AdaRMSNorm timestep conditioning, so the transfer claim is not demonstrated on the stock π0 architecture. Please release code/checkpoints or include a direct comparison on the original backbone to substantiate the comparative claims.
  3. [Section 7 and Appendix D] The noise-substitution discussion contradicts its own table. The text says replacing F/T with Gaussian noise yields 'no statistically significant change on precision-critical tasks including plug insertion and button push,' but Table A.1 shows button push, a force-critical task, dropping from 75.0% to 5.0% (p<.001) under FACT with noise. The text misclassifies button push and conflicts with the table. This ablation is load-bearing for the claim that FACT genuinely exploits force, so the inconsistency must be corrected and the results re-stated consistently.
minor comments (4)
  1. [Section 5.2, Eq. (5)] The notation in Eq. (5) is under-specified: γ_l(τ), β_l(τ), g_l(τ), and the residual connection h_{l-1} are not fully defined. Please clarify the dimensions and the role of g_l, and state whether the residual is added before or after normalization.
  2. [Section 8] The conclusion states that ablations confirm the two components are 'causally independent.' The experiments show additive gains and separate ablations, but no manipulation directly tests causal independence. This is an overstatement; 'complementary' is supported, 'causally independent' is not.
  3. [Tables 2 and 4] Table 2 and Table 4 lack column headers describing which task each success-rate column corresponds to. The text helps, but self-contained headers would improve readability. The reader must infer the plug/key/button ordering from the prose.
  4. [Appendix D] In Table A.1, the p-values for FACT on plug insertion (p=.090) and for ForceVLA on key insertion (p=.006) are reported against unspecified baselines. It would be clearer to state explicitly that Fisher's exact test is computed relative to the real-force condition for each method, rather than against π0.5.

Circularity Check

1 steps flagged · score 5.0 of 10

Precision-failure evidence is partly in-sample: LN location m is tuned on plug insertion, the same task used to claim the precision gain.

  1. fitted input called prediction [Appendix F and Section 7 / Table 1]
    "To justify our choice of m=1.5, we sweep m∈ {−1.5,−0.5,0.5,1.5} at fixed s=1, keeping all other training and evaluation settings identical. All runs use the full FACT model, varying only m. We run this sweep on plug insertion, a precision-critical task ... The π0.5+LN row in Table 1 shows that replacing the Beta schedule with LN improves plug insertion by +20 pp (p=.055) and USB insertion by ∼+10 pp, supporting our diagnosis that the default schedule starves the contact-correction regime on precision-critical tasks."

    The LN location parameter m is selected by sweeping success rate on plug insertion (Appendix F) and picking the value that maximizes performance. Section 7 then cites the plug-insertion improvement as evidence for the precision-failure diagnosis. That improvement is therefore an in-sample fitted outcome, not an independent prediction: the same task used to tune m is the same task used to validate the mechanism. Because m=1.5 was chosen because plug insertion succeeded with it, the +20 pp gain on plug insertion cannot serve as confirmation of the 'training starvation' hypothesis. The USB insertion result is non-significant (p=.249) and does not independently rescue the precision-specific claim. This makes the precision-failure pillar partly circular, although not fully constructed: the LN s

full rationale

The paper is largely self-contained: the force-failure diagnosis is supported by ablations and the noise-substitution experiment, and the time-aware force injection mechanism is not derived from the data it predicts. The LN schedule is motivated by a mathematical property of Beta schedules, and its benefits on force-critical tasks (button push, key insertion) are significant and not part of the m sweep. However, the precision-failure claim rests in part on plug insertion, which is exactly the task used to select the LN location parameter m in Appendix F. That is an in-sample tuned result presented as confirming evidence. No load-bearing self-citation or imported uniqueness theorem appears. The central decomposition is therefore only partially circular; the precision pillar is weakened by fitted evidence, while the force pillar remains independent. Overall score 5 reflects partial circularity.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The central result rests on several hand-chosen hyperparameters, most notably the LN location m=1.5 which is fitted on the headline precision task, and the gradient gating threshold delta=0.5N. The main domain assumptions are about the flow-matching noise regime, the properties of force signals, and the fidelity of re-implemented baselines. No new physical entities are introduced.

free parameters (5)
  • LN location parameter m = 1.5
    Chosen by sweeping m in {-1.5, -0.5, 0.5, 1.5} on plug insertion (Appendix F). The precision-task gains claimed for LN are partly a result of this fitted choice.
  • LN scale parameter s = 1
    Fixed by hand and not swept; noted as 'at fixed s=1' in Appendix F.
  • Gradient gating threshold delta = 0.5 N
    Hand-chosen contact detection threshold for the gradient gate in the force injection (Table A.2). No sensitivity analysis is provided.
  • Force history window H = 30 windows (about 2 s)
    Design choice for temporal context; no ablation over H is reported.
  • Impedance gains kp, kv = Not specified numerically
    Manually tuned controller gains (Appendix C). They affect all methods equally but are part of the system the central result depends on.
assumptions (7)
  • standard math Rectified linear flow-matching interpolant and objective with noise level tau in [0,1]
    Taken from Lipman et al. (2024) and used as the policy training framework in Section 3.
  • domain assumption The Beta(1.5, 1.0) schedule used in pi0.5 allocates 8.9% of gradient signal to tau < 0.2
    Source is the cited pi0 and pi0.5 papers; this number is the basis for the training starvation diagnosis in Section 4.1.
  • domain assumption The low-tau denoising regime governs fine corrective control during contact
    Central to the precision-failure diagnosis in Section 4.1; the paper provides no direct causal verification beyond the LN sweep.
  • domain assumption Force signals exhibit contact sparsity, temporal structure, and sensitivity modulation
    Assumed as the motivation for the force injection design in Section 4.2 and supported only by empirical ablation, not derivation.
  • domain assumption The re-implemented ForceVLA and TA-VLA baselines are faithful to the original methods
    Section 6 states both methods are re-implemented on pi0.5; no verification against original checkpoints or code is provided.
  • domain assumption 40 rollouts per task are sufficient for the reported statistical conclusions
    The paper uses Fisher's exact test on 40 rollouts per cell; no power analysis or confidence intervals are provided.
  • ad hoc to paper The modified pi0 action expert with AdaRMSNorm timestep conditioning preserves representational equivalence to the original pi0
    Appendix E describes this modification to enable timestep conditioning; the paper assumes it does not disadvantage the pi0 backbone comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them." pith.science (2026). https://pith.science/paper/QE2SI5WF

@misc{pith2026260801402,
  author       = {Pith},
  title        = {Pith review of: Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QE2SI5WF}},
  note         = {Machine review of arXiv:2608.01402}
}
read the original abstract

We address the problem of understanding when and why Vision-Language-Action models struggle with contact-rich manipulation tasks that require precise physical interaction. Prior work has primarily focused on addressing contact failures through force-augmented architectures and training-time regularizers, yet the root causes of these failures remain underexplored. We identify two distinct failure modes underlying this gap. Precision failures are rooted in a flow-matching policy training mismatch, and force failures arise from the distinctive structure of force signals. We address each failure mode with a targeted mechanism and combine them into FACT, which achieves 66% average success rate across five contact-rich tasks against 41% for the best prior baseline, in an evaluation spanning almost 2,500 real-world rollouts.

Figures

Figures reproduced from arXiv: 2608.01402 by the authors.

Figure 1
Figure 1. Contact-rich manipulation tasks from our experiments. These tasks involve sustained or repeated physical interaction, where success depends on simultaneously regulating both motion and contact forces measured by a wrist-mounted force/torque sensor. Plug insertion (top) demands sub-millimeter alignment of the connector under partial occlusion while managing contact until the connector is fully seated in the socket. B… view at source ↗
Figure 2
Figure 2. VLA failure modes on plug insertion. Force magnitude |F| is shown over time. Dashed grey traces repeat the success profile for reference. (a) Success. The plug approaches the socket (|F|=0), makes contact and explores alignment (|F|>0), enters the socket (|F|≈0), and fully seats, producing the sharp force rise at t1. (b) Precision failure. The plug is misaligned at the socket entry and force builds and saturates at … view at source ↗
Figure 3
Figure 3. FACT method overview. (a) Logit-Normal (LN) concentrates 6× more gradient signal in the contact-correction regime (τ < 0.2, gray line) than the default Beta(1.5, 1.0) schedule. (b) Time-aware force injection uses the current reading ft to modulate per-layer sensitivity via AdaRM￾SNorm, while the force history ft−H:t−1 is prepended as tokens for temporal context. given as: fT (τ ) = 1 s √ 2π 1 τ (1 − τ ) exp − (logit… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: VLA failures on two contact-rich tasks. Each row shows an approach frame (purple) followed by a success (green) and failure (red) frame with the corresponding force profiles. Vertical lines denote the approach event and the times at which success or failure is register…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 10 canonical work pages

  1. [1]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al.π 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    arXiv 2024

  2. [2]

    Intelligence, K

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al.π 0.5: a vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054, 2025

    arXiv 2025

  3. [3]

    Zitkovich, T

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

    2023

  4. [4]

    Bjorck, F

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

    arXiv 2025

  5. [5]

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

    arXiv 2024

  6. [6]

    Shukor, D

    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 afford- able and efficient robotics.arXiv preprint arXiv:2506.01844, 2025

    arXiv 2025

  7. [7]

    O’Neill, A

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024

    2024

  8. [8]

    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

Show all 27 references
  1. [9]

    Zhang, H

    Z. Zhang, H. Xu, Z. Yang, C. Yue, Z. Lin, H.-a. Gao, Z. Wang, and H. Zhao. TA-VLA: Eluci- dating the design space of torque-aware vision-language-action models. In9th Conference on Robot Learning (CoRL), 2025

    2025

  2. [10]

    Y . Li, P. Tang, W. Zhang, C. Zhu, Y . Duan, W. Shi, X. Zhang, Z. Yang, J. Ji, and Y . Zhang. FA VLA: A force-adaptive fast–slow VLA model for contact-rich robotic manipulation.arXiv preprint arXiv:2602.23648, 2026

    2026

  3. [11]

    Zhang, Y

    Y . Zhang, Y . Wang, X. Sun, K. Huang, Z. Xu, J. Ji, Z. Che, J. Tang, and J. Sun. Craft: Adapting vla models to contact-rich manipulation via force-aware curriculum fine-tuning.arXiv preprint arXiv:2602.12532, 2026

    2026

  4. [12]

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

    2026

  5. [13]

    D. E. Whitney. Quasi-static assembly of compliantly supported rigid parts. 1982. 9

    1982

  6. [14]

    O. Khatib. A unified approach for motion and force control of robot manipulators: The opera- tional space formulation.IEEE Journal on Robotics and Automation, 3(1):43–53, 1987

    1987

  7. [15]

    Lefebvre, H

    T. Lefebvre, H. Bruyninckx, and J. De Schutter. Active compliant motion: A survey.Advanced Robotics, 19(5):479–520, 2005

    2005

  8. [16]

    S. R. Chhatpar and M. S. Branicky. Search strategies for peg-in-hole assemblies with position uncertainty. InProceedings of the 2001 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2001

    2001

  9. [17]

    Noseworthy, B

    M. Noseworthy, B. Tang, B. Wen, A. Handa, C. Kessens, N. Roy, D. Fox, F. Ramos, Y . Narang, and I. Akinola. Forge: Force-guided exploration for robust contact-rich manipulation under uncertainty.IEEE Robotics and Automation Letters, 2025

    2025

  10. [18]

    M. A. Lee, Y . Zhu, P. Zachares, M. Tan, K. Srinivasan, S. Savarese, L. Fei-Fei, A. Garg, and J. Bohg. Making sense of vision and touch: Learning multimodal representations for contact- rich tasks.IEEE Transactions on Robotics, 36(3):582–596, 2020. doi:10.1109/TRO.2019. 2959445

    2020 doi

  11. [19]

    Z. He, H. Fang, J. Chen, H.-S. Fang, and C. Lu. Foar: Force-aware reactive policy for contact- rich robotic manipulation.IEEE Robotics and Automation Letters, 2025

    2025

  12. [20]

    Y . Li, Q. Lyu, J. Yang, Y . Salam, and W. Wang. A hybrid framework using diffusion policy and residual rl for force-sensitive robotic manipulation.IEEE Robotics and Automation Letters, 2025

    2025

  13. [21]

    C. Chen, Z. Yu, H. Choi, M. Cutkosky, and J. Bohg. Dexforce: Extracting force-informed actions from kinesthetic demonstrations for dexterous manipulation.IEEE Robotics and Au- tomation Letters, 2025

    2025

  14. [22]

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

    2025 arXiv

  15. [23]

    J. J. Liu, Y . Li, K. Shaw, T. Tao, R. Salakhutdinov, and D. Pathak. Factr: Force-attending curriculum training for contact-rich policy learning.arXiv preprint arXiv:2502.17432, 2025

    2025 arXiv

  16. [24]

    Lipman, R

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    2022 arXiv

  17. [25]

    Esser, S

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. M¨uller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024

    2024

  18. [26]

    S. Bai, J. Z. Kolter, and V . Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling.arXiv preprint arXiv:1803.01271, 2018

    2018 arXiv

  19. [27]

    Added parameters

    G. Casiez, N. Roussel, and D. V ogel. 1 C filter: a simple speed-based low-pass filter for noisy input in interactive systems. InProceedings of the SIGCHI Conference on human factors in computing systems, pages 2527–2530, 2012. 10 A Task Descriptions We evaluateFACTon the five...

    2012

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.