Pith. sign in

REVIEW 3 major objections 5 minor 37 references

ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read ContactGuard predicts grasp failures before contact by rolling the planned action chunk through a frozen latent world model and vetoing the grasp when the predicted post-contact latent flags failure.

desk verdict Solid, real-robot pre-contact failure monitor with clean ablations; the one missing check is whether the latent rollout actually tracks the post-contact scene. read the letter →

arxiv 2608.13438 v1 pith:WZJXSVNZ submitted 2026-08-13 cs.RO cs.AIcs.CV

classification cs.ROcs.AIcs.CV
keywords pre-contactexecutionmonitoringlatentworldmodelsaction-conditionedpredictiongraspfailurevisuomotorpoliciesJEPAmulti-viewfusion
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

ContactGuard is a pre-contact execution monitor for chunked visuomotor policies. It takes the action chunk the policy is about to execute, rolls it forward through a frozen latent world model trained on unlabelled robot trajectories, and uses a small linear probe on the predicted post-contact latent to decide whether the grasp will fail. The paper's central claim is that this imagined future latent carries failure information not available from the current observation or the planned action alone, and that the monitor can veto a bad grasp before the gripper closes on a live robot without changing the policy. Across four real-world tasks, the predicted future latent improves ROC AUC over a matched current-latent baseline, and action-corruption experiments show the signal is tied to the specific proposed action. The paper is explicit that this is a veto, not a recovery: aborting stops the failure but does not complete the task, and a separate recovery module is left as future work.

What carries the argument

The central object is the action-conditioned latent world model: a shared ViT-Tiny encoder maps each camera view to a compact embedding, the per-view embeddings are mean-pooled across cameras into one latent, and a causal Transformer predictor (conditioned on actions through AdaLN-zero modulation and regularized with SIGReg) is trained with next-latent regression to roll the latent forward under planned actions. This follows the Joint-Embedding Predictive Architecture (JEPA) principle, predicting the embedding of a future signal rather than reconstructing pixels. At deployment the predictor is unrolled autoregressively for 30 steps from an anchor 15 frames before gripper closure, and a frozen ℓ2-regularized logistic regression probe scores the final predicted latent. This machinery converts a concrete proposed action chunk into a predicted visual-consequence latent that a low-capacity linear readout can classify, without pixel-level video prediction.

What would settle it

Collect a labeled set in which the same planned action chunk is executed from the same observation but succeeds and fails due to unobserved object state; if ContactGuard's predicted post-contact latent cannot separate these cases (ROC AUC near 0.5), the imagined-future signal is not carrying the outcome information the paper claims.

Watch

Extended reading notes

Core claim

On its own terms, ContactGuard establishes that an action-conditioned latent world model, trained without any success labels, produces a representation of a planned grasp's near future that a lightweight logistic probe can read to separate successful from failed grasps. Across four real tasks (cup and box pick-and-place, pencil-and-notebook, towel-fold), the predicted future latent achieves higher ROC AUC than the matched current-latent baseline and than external failure detectors, and live robot rollouts show the monitor aborting before closure with lower false-abort rates than direct prediction from the anchor latent and action. The paper also shows the signal is action-specific: holding the observation fixed and swapping in an action chunk from a failed attempt increases the predicted failure probability, while corrupting the action chunk collapses performance to near chance.

Load-bearing premise

The system relies on the compressed visual snapshot averaged across cameras, after being rolled forward by the world model, preserving enough information about contact (slip, displacement, misalignment) to tell success from failure half a second after the gripper closes, even though the world model is never trained on outcome labels.

Editorial extensions

If this is right

  • ContactGuard can be attached to an existing chunked visuomotor policy as a black-box verifier, so deployment does not require retraining or modifying the policy.
  • Pre-contact abort decisions happen within the control loop: the full encode–rollout–probe pass at 30 steps takes about 19 ms, leaving slack before gripper closure at 30 Hz.
  • The monitor's signal comes from the specific pending action, not just a risky-looking scene; corrupting or swapping the action chunk changes the verdict.
  • Multi-view mean-fused latents improve prediction over any single view, particularly when the contact region is occluded from one camera.
  • The monitor is a veto, not a recovery; after an abort the task still needs an external recovery module, which the paper leaves as future work.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The trigger in this paper is a gripper open-to-close transition, but the underlying rollout-and-probe machinery is trigger-agnostic; an equivalent monitor could be built for other imminent contact events such as insertion, pressing, or pushing by only changing the trigger detector.
  • Because the world model is trained without outcome labels, its latent is not optimized to retain slip or misalignment details; a future objective that adds an auxiliary task-outcome or contrastive loss might raise the ceiling, or reveal a failure mode on tasks where those details are not visually expressed.
  • The action-swap result suggests ContactGuard could be used offline as an action-chunk evaluator, screening or ranking candidate chunks before execution rather than only aborting a single chunk at runtime.
  • The false-abort variation across tasks (for example, towel trading precision for coverage) implies threshold choice is consequential; a user deploying this would want to calibrate the threshold on their own success/failure distribution.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces ContactGuard, a pre-contact execution monitor for chunked visuomotor policies. It trains a JEPA-style latent world model on unlabelled robot trajectories to predict multi-view visual latents under planned actions, then trains a linear logistic probe on the predicted post-contact latent to score grasp failure. At deployment, the monitor triggers on an imminent gripper-closure event, rolls the frozen predictor forward under the policy's action chunk, and aborts if the predicted failure probability exceeds a per-task threshold. The paper reports four real-world grasp tasks (cup, box, pencil, towel) and compares against a single-view world model, direct classification from the current latent plus actions, a matched current-latent ablation, and external failure detectors. The main empirical claims are that predicted future latents yield higher ROC AUC than current-latent and direct baselines, that action corruption destroys the signal, and that the system runs fast enough for a 0.5 s pre-contact window. The paper also includes a live real-robot evaluation with abort-based counterfactual outcome recording.

Significance. If the results hold, ContactGuard is a practical contribution: it adds a policy-decoupled predictive verifier to chunked visuomotor policies without retraining or candidate search. The ablation design in Table 2 is a genuine strength: the matched current-latent baseline shares the encoder, probe class, data pools, and cross-validation protocol, and the reported five-seed cross-validation makes the offline comparison credible. The action-swap test in Section 4.3 is a particularly clean falsifiable intervention. However, the paper's mechanistic claim—that the probe reads a faithful forecast of the post-contact latent—is not yet supported, and one baseline is not fully matched on input context. If the missing controls are added and support the current results, this would be a solid systems contribution to runtime monitoring.

major comments (3)
  1. [§3.3, §4.3, Eqs. (3)–(4)] The central mechanistic claim is that the failure signal comes from a faithful prediction of the post-contact latent, but no fidelity check is reported. The paper never compares the autoregressively rolled-out \hat{z}_{t+K} with the true latent z_{t+K} on held-out clips, and it never reports the probe's AUC when fed the true future latent instead of the predicted one. Because the world model is trained only with teacher-forced one-step MSE plus SIGReg (Eq. 4), multi-step rollouts are not guaranteed to stay near the true latent. The current-latent and corrupted-action ablations show that the signal is action-dependent, but a corrupted or unfaithful rollout could still produce a useful action-conditional representation. Add a latent-fidelity measurement (e.g., MSE or cosine similarity between predicted and true future latents at K=1, 15, 30) and a probe-on-real-future-latent control; without these, the abstract's claim that the monitor 'predicts its short-horizon consequence in latent visual space' is not established.
  2. [§4.2, Table 2; §4.3, Table 3] The 'Current latent' ablation is not fully matched on input features. ContactGuard's probe reads \hat{z}_{t+K}, which is a function of the C=3 latent context z_{t-C+1:t} and K future actions, whereas the Current latent baseline reads only the single anchor latent z_t. The observed AUC gain could therefore be due in part to the additional temporal context rather than to the imagined future. Add a baseline that consumes the full context without rollout—for example, the mean or concatenation of z_{t-C+1:t} fed to the same linear probe—to isolate the contribution of the action-conditioned latent prediction.
  3. [§4.2, Table 1] The live closed-loop results in Table 1 are reported as point estimates without any uncertainty quantification. With N=50 per task, the difference between Ours and LeWM on Box (AUC 0.946 vs 0.933, FAR 0.107 vs 0.107) and on Towel (balanced accuracy 0.820 vs 0.700) could be within sampling noise; for example, a single confusion-matrix entry change on Box alters recall by about 0.04. Report bootstrap confidence intervals or repeated-seed statistics for the live metrics, or at minimum state the number of independent rollouts per cell and how thresholds were fixed. This is needed to support the paper's claim that the monitor 'transfers to live robot'.
minor comments (5)
  1. [Table 2, Table 3] Both Table 2 and Table 3 report a 'Current latent' row with different values (e.g., Cup 0.840 vs 0.758) because they use different offline splits. The text says this in passing in §4.3, but the discrepancy should be flagged explicitly in the captions to avoid confusion.
  2. [Figure 1] Figure 1 uses 'h steps' for the rollout horizon while the text and equations use K; unify the notation.
  3. [Abstract] The abstract has several formatting artifacts (e.g., missing space after the ContactGuard emphasis, 'setups' typo); please run a clean proofread and LaTeX pass.
  4. [§4.1] The text states the anchor is 0.5 s before closure and the rollout reaches 0.5 s after closure at 30 Hz, but the phrase could be misread because k_pre=15 frames at 30 Hz is 0.5 s and K=30 steps reaches T_g + 15 frames. Clarify the exact frame indices.
  5. [Eq. (6)] The class-weight notation is slightly confusing: s_c is defined for class c, but the loss uses s_{y_i}; consider defining the weight directly as a function of the sample's class to improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the failure probe is a supervised classifier on predicted latents, the world model is unsupervised, and all central comparisons are matched external baselines.

full rationale

The derivation chain is empirical and self-contained. ContactGuard trains an unsupervised JEPA-style world model with Eq. (4) on unlabelled trajectories, freezes it, and then trains a supervised logistic probe (Eq. (6)) on labelled pre-contact clips using the predicted latent \hat{z}_{t+K} as the feature. No parameter is fitted to test labels; the probe and per-task threshold are fixed on train/validation splits and evaluated on held-out live rollouts and offline pools. The central comparisons against 'Current latent' and corrupted-action ablations are matched interventions that differ only in whether the probe reads z_t or the rolled-out \hat{z}_{t+K}; the action-swap test holds the observation fixed while swapping the action chunk, showing the signal depends on the proposed action rather than static visual risk. The paper does not invoke any uniqueness theorem, and the related-work citations to Rewind-IL, PATCH, and Dose3 are not load-bearing for the empirical claim. The skeptic's concern that multi-step rollout fidelity to the true post-contact latent is never directly measured is a correctness/interpretability limitation, not a circular reduction: the paper never defines success as matching the true latent, and the reported ROC AUC is an external benchmark against matched baselines. No equation reduces to its own input, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The world model is trained on unlabelled trajectories, which keeps its supervision cost low, but the method's success rests on several domain assumptions: that gripper closure is a reliable anchor, that the mean-pooled multi-view latent preserves outcome-relevant contact physics, that the world model generalizes to failed grasps, and that a linear probe is sufficient. These assumptions are reasonable but not directly validated.

free parameters (3)
  • Per-task abort threshold tau = Cup 0.605; Box 0.487; Pencil 0.580; Towel 0.415
    Selected on validation split per task (Section 3.5), then frozen. The reported false-abort rates and operating points depend on these values.
  • Anchor offset k_pre and rollout horizon K = k_pre = 15 frames (0.5 s before closure); K = 30 steps (1 s, readout 0.5 s after closure)
    Hand-chosen detection window (Sections 3.4 and 3.5, Appendix A). These set how early the monitor acts and which future latent is read; performance is likely sensitive to them.
  • Probe regularization rho and class-balance weights = rho = 1; s_c = N/(2 N_c)
    Logistic-regression settings (Eq. 6). The class weights are derived from label counts; rho is fixed. Minor influence on the decision boundary.
assumptions (5)
  • domain assumption The gripper open-to-close transition in the commanded action reliably marks imminent contact and defines the anchor time T_g.
    Section 3.5 uses this command-based trigger as the pre-contact anchor. If closure is not imminent or the command is ambiguous, the monitor cannot activate.
  • domain assumption The mean-pooled multi-view latent preserves the physical details (slip, displacement, misalignment) needed to infer grasp outcome after roll-out.
    Section 3.3 mean-pools per-view embeddings; Section 3.4 reads out the predicted latent. No training objective explicitly enforces retention of task-outcome details.
  • domain assumption A world model trained with next-latent MSE and SIGReg on unlabelled trajectories generalizes to the deployment distribution, including failed grasp attempts.
    Section 3.2 and Appendix A. The unlabelled data come from ACT rollouts and teleoperation; the coverage of failures is not quantified.
  • domain assumption A linear logistic probe on the standardized predicted latent is sufficiently expressive to separate success and failure.
    Section 3.4. Nonlinear probes on the anchor latent and action are unstable, but nonlinear probes on the predicted latent are not tested.
  • standard math Standard statistical tools (logistic regression, bootstrap, cross-validation) are valid for the reported comparisons.
    Used for probe fitting and uncertainty estimates in Table 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models." pith.science (2026). https://pith.science/paper/WZJXSVNZ

@misc{pith2026260813438,
  author       = {Pith},
  title        = {Pith review of: ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WZJXSVNZ}},
  note         = {Machine review of arXiv:2608.13438}
}
read the original abstract

Contact-rich manipulation failures are often detected only after the robot has committed to contact. This is especially limiting in wrist-camera setups: close gripper--object views help observe contact, but a poor approach may already push, miss, slip, or disturb the object before conventional detectors react. We introduce \emph{ContactGuard}, a pre-contact execution monitor for chunked visuomotor policies. Given the policy's planned action chunk, ContactGuard predicts its short-horizon consequence in latent visual space and aborts if the predicted future latent indicates likely failure. Its latent world model is trained from unlabelled robot trajectories to predict compact multi-view visual embeddings under planned actions, avoiding pixel-level video prediction. A lightweight failure probe is then trained from a small labelled set of pre-contact clips. At deployment, ContactGuard anchors prediction before an imminent contact event, rolls the model forward under the policy's own actions, and verifies the predicted post-contact latent. Across real-world contact-rich manipulation tasks, ContactGuard predicts failure more accurately than direct and corrupted-action ablations, and transfers to live robot as a pre-contact abort signal without modifying the underlying policy.

Figures

Figures reproduced from arXiv: 2608.13438 by the authors.

Figure 1
Figure 1. Pre-contact online monitoring via action-conditioned latent world models. While a visuomotor policy executes an action chunk, ContactGuard conditions a frozen latent world model on the same planned chunk at an anchor k steps before the planned gripper closure Tg. It rolls the model forward h steps and scores the predicted future latent with a lightweight classifier. If P(fail) > τ , execution is aborted before conta… view at source ↗
Figure 2
Figure 2. ContactGuard overview. A shared multi-view encoder maps camera observations to a latent context. An action-conditioned predictor rolls this context forward under the policy’s planned action chunk to produce future latents zˆt+1:t+K. A lightweight probe is trained offline on labelled pre-contact clips and used online to score the predicted post-contact latent zˆt+K. If the predicted failure probability exceeds thresh… view at source ↗
Figure 3
Figure 3. Evaluated tasks. Each panel shows a representative successful rollout from the middle [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Real-robot qualitative examples on the pick-and-place tasks. Each panel shows a live [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative examples on the pencil-and-notebook and towel-fold tasks. Each panel shows a rollout that is classified correctly [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 16 canonical work pages

  1. [1]

    Assran, Q

    M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y . LeCun, and N. Bal- las. Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023

  2. [2]

    Ha and J

    D. Ha and J. Schmidhuber. World Models. Mar. 2018. arXiv:1803.10122

  3. [3]

    Hafner, T

    D. Hafner, T. Lillicrap, I. Fischer, R. Villegas, D. Ha, H. Lee, and J. Davidson. Learning Latent Dynamics for Planning from Pixels, June 2019. arXiv:1811.04551

  4. [4]

    Hafner, T

    D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi. Dream to Control: Learning Behaviors by Latent Imagination, Mar. 2020. arXiv:1912.01603

  5. [5]

    Hafner, J

    D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap. Mastering Diverse Domains through World Models, Apr. 2024. arXiv:2301.04104

  6. [6]

    Hansen, X

    N. Hansen, X. Wang, and H. Su. Temporal Difference Learning for Model Predictive Control, July 2022. arXiv:2203.04955 [cs.LG]

  7. [7]

    Srinivas, A

    A. Srinivas, A. Jabri, P. Abbeel, S. Levine, and C. Finn. Universal Planning Networks: Learn- ing Generalizable Representations for Visuomotor Control. InProceedings of the 35th Inter- national Conference on Machine Learning, pages 4732–4741. PMLR, July 2018

  8. [8]

    Finn and S

    C. Finn and S. Levine. Deep Visual Foresight for Planning Robot Motion, Mar. 2017. arXiv:1610.00696 [cs.LG]

Show all 37 references
  1. [9]

    F. Zhu, H. Wu, S. Guo, Y . Liu, C. Cheang, and T. Kong. IRASim: A Fine-Grained World Model for Robot Manipulation, July 2025. arXiv:2406.14540 [cs.RO]

  2. [10]

    Barcellona, A

    L. Barcellona, A. Zadaianchuk, D. Allegro, S. Papa, S. Ghidoni, and E. Gavves. Dream to Manipulate: Compositional World Models Empowering Robot Imitation Learning with Imag- ination, Mar. 2025. arXiv:2412.14957 [cs.RO]

  3. [11]

    L. Maes, Q. L. Lidec, D. Scieur, Y . LeCun, and R. Balestriero. LeWorldModel: Stable End- to-End Joint-Embedding Predictive Architecture from Pixels, Mar. 2026. arXiv:2603.19312 [cs.LG]

  4. [12]

    Y . Du, K. Konyushkova, M. Denil, A. Raju, J. Landon, F. Hill, N. d. Freitas, and S. Cabi. Vision-Language Models as Success Detectors, Mar. 2023. arXiv:2303.07280

  5. [13]

    J. Duan, W. Pumacay, N. Kumar, Y . R. Wang, S. Tian, W. Yuan, R. Krishna, D. Fox, A. Man- dlekar, and Y . Guo. AHA: A Vision-Language-Model for Detecting and Reasoning Over Fail- ures in Robotic Manipulation, Oct. 2024. arXiv:2410.00371

  6. [14]

    Liang, Y

    A. Liang, Y . Korkmaz, J. Zhang, M. Hwang, A. Anwar, S. Kaushik, A. Shah, A. S. Huang, L. Zettlemoyer, D. Fox, Y . Xiang, A. Li, A. Bobu, A. Gupta, S. Tu, E. Biyik, and J. Zhang. Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons, May 2026

  7. [15]

    T. Lee, A. Wagenmaker, K. Pertsch, P. Liang, S. Levine, and C. Finn. RoboReward: General- Purpose Vision-Language Reward Models for Robotics, Jan. 2026. arXiv:2601.00675

  8. [16]

    Cheng, T

    H. Cheng, T. Zheng, Z. Ma, T. Zhang, M. Johnson-Roberson, and W. Zhi. Dose3: Diffusion- based unified out-of-distribution detection onSE(3)trajectories.IEEE Robotics and Automa- tion Letters, 11(2), 2026

  9. [17]

    R ¨omer, A

    R. R ¨omer, A. Kobras, L. Worbis, and A. P. Schoellig. Failure Prediction at Runtime for Gen- erative Robot Policies, Oct. 2025. arXiv:2510.09459. 10

  10. [18]

    C. Xu, T. K. Nguyen, E. Dixon, C. Rodriguez, P. Miller, R. Lee, P. Shah, R. Ambrus, H. Nishimura, and M. Itkina. Can we detect failures without failure data? InRSS, 2025

  11. [19]

    Burda, H

    Y . Burda, H. Edwards, A. Storkey, and O. Klimov. Exploration by random network distillation. InICLR, 2019

  12. [20]

    Q. Gu, Y . Ju, S. Sun, I. Gilitschenski, H. Nishimura, M. Itkina, and F. Shkurti. SAFE: Multitask failure detection for vision-language-action models. InNeurIPS, 2025

  13. [21]

    Zheng, S

    G. Zheng, S. Seenivasan, M. Johnson-Roberson, and W. Zhi. Rewind-il: Online failure detec- tion and state respawning for imitation learning.arXiv preprint arXiv:2604.16683, 2026

  14. [22]

    Y . Zhou, R. Qiu, Y . Chen, J. Cui, and W. Zhi. Patch: Action-chunk-conditioned latent patch innovation monitoring for robot manipulation.arXiv preprint arXiv:2606.16690, 2026

  15. [23]

    H. Liu, S. Dass, R. Mart ´ın-Mart´ın, and Y . Zhu. Model-based runtime monitoring with interac- tive il. InICRA, 2024

  16. [24]

    H. Liu, Y . Zhang, V . Betala, E. Zhang, J. Liu, C. Ding, and Y . Zhu. Multi-task interactive robot fleet learning with visual world models. InCoRL, 2024

  17. [25]

    Damak, M

    K. Damak, M. Boujelbene, C. Acun, A. Alvanpour, S. K. Das, D. O. Popa, and O. Nasraoui. Robot failure mode prediction with deep learning sequence models.Neural Computing and Applications, 37:4291–4302, Feb. 2025

  18. [26]

    W. Zhi, T. Lai, L. Ott, and F. Ramos. Diffeomorphic transforms for generalised imitation learning. InLearning for Dynamics and Control Conference, L4DC, 2022

  19. [27]

    Ravichandar, A

    H. Ravichandar, A. S. Polydoros, S. Chernova, and A. Billard. Recent advances in robot learning from demonstration.Annual review of control, robotics, and autonomous systems, 2020

  20. [28]

    W. Zhi, H. Tang, T. Zhang, and M. Johnson-Roberson. Teaching periodic stable robot motion generation via sketch.IEEE Robotics and Automation Letters, 2025

  21. [29]

    W. Zhi, I. Akinola, K. van Wyk, N. Ratliff, and F. Ramos. Global and reactive motion genera- tion with geometric fabric command sequences. InIEEE International Conference on Robotics and Automation, ICRA. IEEE, 2023

  22. [30]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware, Apr. 2023. arXiv:2304.13705

  23. [31]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion, Mar. 2024. arXiv:2303.04137

  24. [32]

    Z. Li, Y . Zhou, R. Qiu, H. Wu, G. Ren, and W. Zhi. Tripilot-ff: Coordinated whole-body teleoperation with force feedback.arXiv preprint arXiv:2602.09888, 2026

  25. [33]

    M. Dai, L. Liu, Y . Bai, Y . Liu, Z. Wang, R. SU, C. Chen, L. Lin, and X. Wu. RoVer: Robot Reward Model as Test-Time Verifier for Vision-Language-Action Model, Oct. 2025. arXiv:2510.10975 [cs.RO]

  26. [34]

    X. Sun, Z. Xu, C. Cao, Z. Liu, Y . Sun, J. Pang, R. Zhang, Z. Yang, K. Pang, D. He, M. Yuan, and J. Chen. AtomVLA: Scalable Post-Training for Robotic Manipulation via Predictive Latent World Models, Mar. 2026. arXiv:2603.08519 [cs.RO]

  27. [35]

    Tang and W

    J. Tang and W. Zhi. Autointervene: Calibrated intervention for action-chunking imitation learning policies.arXiv preprint arXiv:2608.07065, 2026

  28. [36]

    Lipman, R

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations, 2023. 11

  29. [37]

    small” and “large

    J. Long, D. Liu, W. Cai, I. Manchester, and W. Zhi. Safe policies post-training: Constraining streaming flow models for adapting learned robot trajectory distributions.IEEE Robotics and Automation Letters, 11(9), 2026. 12 A Implementation Details Encoder and latent dimensions....

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.