Pith. sign in

REVIEW 3 major objections 3 minor 44 references

Track4Action: Distilling World-Centric 3D Tracker into Vision-Language-Action Policies

T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read A current-observation VLA policy can be trained to predict the 3D world transition of its own action chunk from a frozen tracker, then deployed without the tracker or clip.

desk verdict A well-scoped tracker-distillation idea with broad experiments; the central causal claim outruns the only ablation, which removes the whole query-gate pathway along with the alignment loss. read the letter →

arxiv 2608.03727 v1 pith:6JGFNJKZ submitted 2026-08-04 cs.RO

classification cs.RO
keywords vision-language-actionpolicies3Dtrackingprivilegedsupervisionfeaturedistillationflowmatchingrobotmanipulationbimanualzero-shotrobustness
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

Demonstration action labels tell a robot policy which commands to execute but not how those commands change the 3D scene. Track4Action's claim is that the missing supervision can be supplied by a frozen world-centric 3D tracker: during training, the tracker reads the demonstration clip whose K frame transitions match the K actions, and the policy is trained to infer the pooled tracker feature from its current observation and language instruction alone. The same learned track queries that match this target also condition a flow-matching action head through a feature-wise gate, so the world-transition objective shapes the representation that actually predicts actions, while neither the clip nor the tracker is present at deployment. The method reports 97.0% average success on LIBERO, 82.3% on zero-shot LIBERO-Plus (7.6 points over its alignment-free control and 3.0 over LaMP), 80.44%/81.48% on RoboTwin 2.0, and 67.5% average success on four physical bimanual tasks (25.0 points over the control). If true, action-aligned 3D tracker features are an effective form of privileged supervision for robust, tracker-free VLA control.

What carries the argument

The carry mechanism is the action-aligned track-query pathway. A frozen Track4World teacher pools tokens over the K frame transitions aligned with a K-action chunk into a target feature; learnable track queries cross-attend to the current VLA hidden states to form a student representation; L2 alignment in a shared space trains the queries to infer that world transition from current context; and a feature-wise gate fuses the same queries back into the VLA sequence so the flow-matching action head consumes the aligned representation. This construction ties the privileged target to control while leaving deployment tracker-free.

What would settle it

Run the exact same training with the track-query and gated-fusion pathway intact but the alignment loss set to zero (or with a randomly shuffled tracker target); if LIBERO-Plus and physical success stay near 82.3% and 67.5%, the gains come from added pathway capacity, not from the world-centric tracker target. A complementary check: keep the alignment loss but drop the queries from the action head, testing whether the action-facing fusion is what makes the target matter.

Watch

Extended reading notes

Core claim

The paper establishes that a VLA policy's action-facing representation can be shaped by the exact realized world transition of each action chunk, using a frozen world-centric 3D tracker as teacher. Track4World encodes the aligned clip V_{t:t+K}, pooling its scene, motion, visibility, and camera tokens into a target feature; learnable track queries read current VLA hidden states and are pulled to this target in a shared alignment space by an L2 loss. Because the same queries are fused back into the policy sequence to condition the flow-matching action head, the alignment objective is not a detached auxiliary task but supervision on the representation that generates actions. At evaluation, the

Load-bearing premise

The central comparison removes the track-query and gated-fusion pathway together with the alignment loss, so crediting the gains to tracker supervision assumes this system-level ablation isolates the teacher target's contribution.

Editorial extensions

If this is right

  • A VLA can be given world-transition supervision without changing its deployment interface or adding any runtime dependency on video or tracking models.
  • The same distillation interface transfers across observation counts, action spaces, embodiments, and horizons: single-arm delta actions on LIBERO, three-view 14-D bimanual joint commands on RoboTwin 2.0, and physical bimanual manipulation.
  • Tracker alignment improves success under all seven LIBERO-Plus zero-shot perturbations, with the largest gains on camera changes, noise, and robot variations, indicating robustness rather than a single-category advantage.
  • The gains are largest where world state and multi-stage continuity matter: LIBERO-Long improves 9.6 points over the control, and physical multi-stage tasks gain 20 to 30 points per task.
  • Because the teacher is frozen and offline, the approach adds no inference cost and is compatible with arbitrary VLA backbones and flow-matching action heads.

Reading between the lines

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

  • The system-level ablation means the reported gains could come partly from the extra track-query and gated-fusion pathway rather than the tracker target; keeping those components but disabling the alignment loss (or aligning to a shuffled target) would separate capacity from supervision.
  • The same recipe should work with other motion-specialized teachers such as point-trackers, optical-flow models, or 4D reconstructions, so the principle may be general privileged-dynamics distillation rather than a Track4World-specific effect.
  • Because the pooled target discards local contact and per-point correspondence, finer-grained alignment targets could extend the gains to deformable manipulation, where the paper's physical results are weakest.
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 / 3 minor

Summary. The paper proposes Track4Action, a method for distilling a frozen world-centric 3D tracker (Track4World) into a current-observation VLA policy. During training, the tracker encodes the K-frame demonstration clip aligned with the action chunk into a pooled feature; learnable track queries read VLA hidden states, are aligned to this feature via an MSE loss in a shared space, and condition a flow-matching action head through gated fusion. At deployment the tracker and clip are removed. The method is evaluated on LIBERO, zero-shot LIBERO-Plus, RoboTwin 2.0, and physical bimanual tasks, reporting improvements over an 'alignment-free' variant and prior baselines.

Significance. The framework is cleanly motivated and the breadth of evaluation (zero-shot robustness, 50-task bimanual simulation, physical experiments) is a strength. The teacher is external (Track4World) and deployment is genuinely tracker-free; the training objective (Eq. 7) is straightforward. If the attribution were supported, the paper would make a useful contribution to privileged supervision for VLA policies. However, the central causal claim is not supported by the reported ablation, which removes the entire query/gating pathway together with the alignment loss; the reported gains could be due to added capacity or inductive bias. The stress-test concern is therefore validated.

major comments (3)
  1. [§4.5, Eq. (6)–(7)] The 'w/o Align' control removes the track-query cross-attention (Eq. 5), the gated fusion into H_cond (Sec. 3.4), and L_align simultaneously. The paper explicitly calls this a 'system-level ablation of the complete tracker-distillation pathway.' As a result, the 7.6-point LIBERO-Plus gain, the 39–42-point RoboTwin gains, and the 25-point physical gains cannot be attributed specifically to tracker supervision. The control lacks the entire extra capacity and representation pathway. A matched ablation is required: keep the query-gate pathway and remove only L_align (or replace the tracker target with a non-informative target) to isolate the effect of the tracker-supervision target. Until then, the abstract's conclusion that the gains 'support action-aligned 3D tracker features as privileged supervision' is unsupported.
  2. [Table 3, §4.4] The alignment-free control scores 41.32/39.90 on RoboTwin, close to the π0.5 level, while Track4Action scores 80.44/81.48. This massive gap is implausibly large for an auxiliary regularizer alone and more consistent with the added cross-attention/gating pathway providing necessary capacity for bimanual action generation. The paper should report the same-architecture no-alignment baseline (as above) to disentangle these factors; without it, the RoboTwin result is not evidence for tracker distillation.
  3. [§5] Physical results are based on 10 trials per task with no confidence intervals or multiple seeds. For example, towel folding is 40% for Track4Action vs 70% for π0.5, a 3-trial difference on 10 trials. The claimed '25.0 points above alignment-free' on physical success and the OOD average (23.3%→50.0%) may be within sampling noise. Report per-trial counts, confidence intervals, and repeated-seed simulation results to support the cross-benchmark claims.
minor comments (3)
  1. [Eq. (6)] The norm notation in the alignment loss appears malformed: '\bar z_t - \bar f_t^{trk}\|^2_2' is missing an opening delimiter. Please fix for clarity.
  2. [§3.1] The definition of V_{t:t+K} and the correspondence between K actions and K transitions is slightly ambiguous. State explicitly that the clip contains K+1 frames and that action i produces the transition from frame t+i-1 to t+i.
  3. [Tables 1–3] The reported averages would be easier to judge with standard errors or confidence intervals, especially for RoboTwin per-task scores where many entries are near 0 or 100.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular reduction: tracker target is external (Track4World) and L_align is a regularizer; the only same-group citation (LaMP) is a baseline, not a load-bearing derivation step.

full rationale

The derivation chain is not circular. The teacher target f_trk is produced by a frozen external tracker (Eq. 3, Track4World [23]) from the action-aligned clip; it is not defined in terms of the action loss or the student outputs. The student track queries are trained by L_align (Eq. 6), an auxiliary regularizer that pulls pooled queries toward this external target, while the action head is trained by a standard flow-matching loss L_act; the full objective (Eq. 7) is a weighted sum of two distinct terms with no term reducing to the other. The paper explicitly keeps the teacher feature out of the action head ('The teacher feature... never enters H_cond' in Sec. 3.4) and removes clip and tracker at deployment. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, and no ansatz is justified by a self-citation. The only same-group citation is LaMP [35], used as a baseline comparison, not as load-bearing support for the method. The 'w/o Align' control (Sec. 4.5) is a system-level ablation that removes the complete tracker-distillation pathway (query cross-attention + gated fusion + L_align), so it cannot isolate L_align's contribution; that is a soundness/attribution limitation, not a circularity, because the comparison is empirical and the target itself remains external. The Limitations section (Sec. 6) further acknowledges the method's dependence on the order of demonstrations and inherited tracker errors, none of which indicate that a result is assumed in its own derivation. Therefore no reduction of the claimed prediction to the input is exhibited.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

The central claim rests on the frozen tracker representation, the temporal alignment of clip and action chunk, the learnability of the query bottleneck, and the validity of the alignment-free control. The control is the most fragile piece: it removes the entire added pathway, so the causal role of the tracker target is not isolated.

free parameters (3)
  • Alignment loss weight λ_align = unspecified
    Weights L_align relative to L_act; not reported in the paper; likely tuned per benchmark; central to the method's effect.
  • Number of track queries and shared-space dimension = unspecified
    Capacity of the bottleneck is not reported; affects whether the queries can express the tracker feature.
  • Pooling scheme over scene/motion/camera tokens = mean pooling (implied)
    The paper says 'pooled' but does not ablate the pooling strategy; a hand-chosen reduction.
assumptions (5)
  • domain assumption Frozen Track4World reliably estimates world-centric 3D motion from the primary-view clip
    The teacher features are trusted as supervision (Section 3.2); limitations admit errors under occlusion and fast motion.
  • domain assumption Exact temporal alignment between K actions and K frame transitions
    Requires that the clip frames correspond one-to-one with the action chunk (Section 3.1); otherwise the target would be mismatched to the supervised commands.
  • domain assumption Track queries can learn to infer the pooled tracker feature from current observations
    The entire distillation relies on this learnability; the paper gives no probing analysis of what the queries encode.
  • ad hoc to paper The alignment-free variant is a matched control that isolates the tracker-supervision effect
    The control removes the whole query/gate pathway plus the alignment loss, conflating architecture and supervision; this is a questionable premise of the experimental logic.
  • standard math Flow-matching objective is a standard valid action-generation loss
    Used from Lipman et al. [21]; not in dispute.
invented entities (2)
  • Learnable track queries Q
    purpose: Extract a compact representation from VLA hidden states via cross-attention, match the tracker feature, and condition the action head through a feature-wise gate
    New architectural component introduced by this paper; no falsifiable handle outside the paper; necessity shown only through internal ablations.
  • Feature-wise gate
    purpose: Fuses the track queries with the VLA sequence to form the action condition Hcond
    A learned component; no independent evidence of its necessity beyond this paper's experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Track4Action: Distilling World-Centric 3D Tracker into Vision-Language-Action Policies." pith.science (2026). https://pith.science/paper/6JGFNJKZ

@misc{pith2026260803727,
  author       = {Pith},
  title        = {Pith review of: Track4Action: Distilling World-Centric 3D Tracker into Vision-Language-Action Policies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6JGFNJKZ}},
  note         = {Machine review of arXiv:2608.03727}
}
abstract

Action labels tell a vision-language-action (VLA) policy which robot commands to imitate, but not how those commands change the 3D world. The aligned demonstration clip contains this missing supervision because its $K$ frame transitions record the geometry, motion, visibility, and camera change produced during the corresponding $K$ actions. We introduce Track4Action, a framework that distills this realized transition from a frozen world-centric 3D tracker into a current-observation VLA policy. During training, Track4World encodes the clip $V_{t:t+K}$ into a pooled tracker feature. Learnable track queries infer this feature from current VLA hidden states, match it in a shared space, and condition a flow-matching action head through a feature-wise gate. The tracker feature only defines the alignment target, so neither the clip nor the tracker is used at deployment. Track4Action reaches 82.3% on zero-shot LIBERO-Plus, improving the alignment-free variant by 7.6 points and LaMP by 3.0 points. It obtains 80.44% and 81.48% on the clean and randomized RoboTwin 2.0 splits, and 67.5% average success across four physical bimanual tasks, 25.0 points above the alignment-free variant. The gains across simulation and physical tasks support action-aligned 3D tracker features as privileged supervision for tracker-free VLA deployment. Our project page is available at https://wing0night.github.io/track4action-project-page.

Figures

Figures reproduced from arXiv: 2608.03727 by the authors.

Figure 1
Figure 1. Track4Action: Unifying 4D Scene Dynam￾ics with Task Understanding. Flow-tracking models cap￾ture how a scene evolves in 4D but lack task awareness, whereas conventional VLA policies understand what to ac￾complish but do not explicitly model future 4D dynamics. Track4Action bridges these complementary strengths, learn￾ing task-conditioned actions with awareness of scene flow. features [28, 33, 36], while geometry-awa… view at source ↗
Figure 2
Figure 2. World-centric 3D tracker feature distillation in Track4Action. During training, frozen Track4World encodes the action-aligned clip Vt:t+K into a pooled target built from scene, motion, and camera tokens. Track queries extracted from current VLA hidden states match this target and are fused back into the VLA sequence through a feature-wise gate to condition the flow-matching action head. The K actions correspond to t… view at source ↗
Figure 3
Figure 3. Illustration of Real-world evaluation. Upper left: the AgileX ALOHA-style hardware platform, built with two Piper arms and equipped with three cameras: one front-facing camera and one wrist-mounted camera on each arm. Center: four real-world manipulation tasks covering rigid-object transfer, deformable-object manipulation, and long-horizon interaction. Right: the OOD settings, including unseen towel color, object la… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Real-world task success rates. Success rate mea￾sures full task completion over 10 trials per task. Across the four tasks, Track4Action averages 67.5% success. The “w/o alignment” legend denotes the variant trained without tracker supervision [PITH_FULL_IMAGE:figures/…
Figure 5
Figure 5. Figure 5: Four-stage process scores on the real-world tasks. Each stacked bar decomposes the process score into four ordered, equally weighted milestones; its total height is the resulting score out of 100. Across the four tasks, Track4Action averages 75.0. to unseen background …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 21 canonical work pages

  1. [1]

    arXiv:2405.01527

    Bharadhwaj,H.;Mottaghi,R.;Gupta,A.;andTulsiani, S.2024.Track2Act:PredictingPointTracksfromInter- netVideosEnablesGeneralizableRobotManipulation. arXiv:2405.01527

  2. [2]

    Bi, H.; Tan, H.; Xie, S.; Wang, Z.; Huang, S.; Liu, H.; Zhao,R.;Feng,Y.;Xiang,C.;Rong,Y.;Zhao,H.;Liu, H.; Su, Z.; Ma, L.; Su, H.; and Zhu, J. 2026. Motus: AUnifiedLatentActionWorldModel. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 35101–35113

  3. [3]

    Y.; Ghosh, D.; Groom, L.; Hausman, K.; Ichter, B.; Jakubczak, S.; Jones, T.; Ke, L.; LeBlanc, D.; Levine, S.; Li-Bell, A.; Mothukuri, M.; Nair, S.; Pertsch, K.; Ren, A

    Black, K.; Brown, N.; Darpinian, J.; Dhabalia, K.; Driess,D.;Esmail,A.;Equi,M.R.;Finn,C.;Fusai,N.; Galliker, M. Y.; Ghosh, D.; Groom, L.; Hausman, K.; Ichter, B.; Jakubczak, S.; Jones, T.; Ke, L.; LeBlanc, D.; Levine, S.; Li-Bell, A.; Mothukuri, M.; Nair, S.; Pertsch, K.; Ren, A. Z.; Shi, L. X.; Smith, L.; Sprin- genberg, J. T.; Stachowicz, K.; Tanner, J....

  4. [4]

    InRobotics: Science and Systems XXI

    Black,K.;Brown,N.;Driess,D.;Esmail,A.;Equi,M.; Finn, C.; Fusai, N.; Groom, L.; Hausman, K.; Ichter, B.; Jakubczak, S.; Jones, T.; Ke, L.; Levine, S.; Li- Bell, A.; Mothukuri, M.; Nair, S.; Pertsch, K.; Shi, L.; Smith,L.;Tanner,J.;Vuong,Q.;Walling,A.;Wang,H.; andZhilinsky,U.2025.π 0:AVision-Language-Action Flow Model for General Robot Control. InRobotics: ...

  5. [5]

    Brohan,A.;Brown,N.;Carbajal,J.;Chebotar,Y.;Chen, X.; Choromanski, K.; Ding, T.; Driess, D.; Dubey, A.; Finn,C.;Florence,P.;Fu,C.;Arenas,M.G.;Gopalakr- ishnan,K.;Han,K.;Hausman,K.;Herzog,A.;Hsu,J.; Ichter,B.;Irpan,A.;Joshi,N.;Julian,R.;Kalashnikov, D.; Kuang, Y.; Leal, I.; Lee, L.; Lee, T.-W. E.; Levine, S.; Lu, Y.; Michalewski, H.; Mordatch, I.; Pertsch, ...

  6. [6]

    Brohan, A.; Brown, N.; Carbajal, J.; Chebotar, Y.; Dabis, J.; Finn, C.; Gopalakrishnan, K.; Hausman, K.; Herzog,A.;Hsu,J.;Ibarz,J.;Ichter,B.;Irpan,A.;Jack- son,T.;Jesmonth,S.;Joshi,N.;Julian,R.;Kalashnikov, D.; Kuang, Y.; Leal, I.; Lee, K.-H.; Levine, S.; Lu, Y.; Malla, U.; Manjunath, D.; Mordatch, I.; Nachum, O.; Parada,C.;Peralta,J.;Perez,E.;Pertsch,K.;...

  7. [7]

    Bu, Q.; Yang, Y.; Cai, J.; Gao, S.; Ren, G.; Yao, M.; Luo, P.; and Li, H. 2025. Learning to Act Anywhere withTask-CentricLatentActions. InRobotics:Science and Systems

  8. [8]

    World- VLA: Towards Autoregressive Action World Model

    Cen,J.;Yu,C.;Yuan,H.;Jiang,Y.;Huang,S.;Guo,J.; Li,X.;Song,Y.;Luo,H.;Wang,F.;etal.2025. World- VLA: Towards Autoregressive Action World Model. arXiv:2506.21539

Show all 44 references
  1. [9]

    Chen, T.; Chen, Z.; Chen, B.; Cai, Z.; Liu, Y.; Li, Z.; Liang, Q.; Lin, X.; Ge, Y.; Gu, Z.; Deng, W.; Guo, Y.; Nian, T.; Xie, X.; Chen, Q.; Su, K.; Xu, T.; Liu, G.; Hu, M.; Gao, H.-a.; Wang, K.; Liang, Z.; Qin, Y.; Yang, X.; Luo, P.; and Mu, Y. 2025. RoboTwin 2.0: A Scalable D...

  2. [10]

    Chen, X.; Wei, H.; Zhang, P.; Zhang, C.; Wang, K.; Guo, Y.; Yang, R.; Wang, Y.; Xiao, X.; Zhao, L.; Chen, J.; and Bian, J. 2025. villa-X: Enhancing Latent Action Modeling in Vision-Language-Action Models. arXiv:2507.23682

  3. [11]

    Doersch, C.; Yang, Y.; Vecerik, M.; Gokay, D.; Gupta, A.; Aytar, Y.; Carreira, J.; and Zisserman, A. 2023. TAPIR: Tracking Any Point with per-frame Initializa- tion and temporal Refinement. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, 10027–10038

  4. [12]

    Fei,S.;Wang,S.;Shi,J.;Dai,Z.;Cai,J.;Qian,P.;Ji,L.; He,X.;Zhang,S.;Fei,Z.;etal.2025.LIBERO-Plus:In- DepthRobustnessAnalysisofVision-Language-Action Models.arXiv preprint arXiv:2510.13626

  5. [13]

    Ghosh,D.;Walke,H.;Pertsch,K.;Black,K.;Mees,O.; Dasari,S.;Hejna,J.;Kreiman,T.;Xu,C.;Luo,J.;Tan, Y.; Chen, L.; Vuong, Q.; Xiao, T.; Sanketi, P.; Sadigh, D.; Finn, C.; and Levine, S. 2024. Octo: An Open- Source Generalist Robot Policy. InRobotics: Science and Systems XX

  6. [14]

    Guan, J.; Zhao, W.; Pei, Y.; Chen, Z.; Solin, A.; and Kannala, J. 2026. Point Tracking Improves World Ac- tion Models. arXiv:2605.23856

  7. [15]

    Karaev, N.; Rocco, I.; Graham, B.; Neverova, N.; Vedaldi, A.; and Rupprecht, C. 2024. CoTracker: It Is Better to Track Together. InComputer Vision – ECCV 2024, 18–35

  8. [16]

    J.; Finn, C.; and Liang, P

    Kim, M. J.; Finn, C.; and Liang, P. 2025. Fine- Tuning Vision-Language-Action Models: Optimizing Speed and Success. arXiv:2502.19645

  9. [17]

    Kim, M. J.; Pertsch, K.; Karamcheti, S.; Xiao, T.; Bal- akrishna, A.; Nair, S.; Rafailov, R.; Foster, E.; Lam, G.; Sanketi, P.; Vuong, Q.; Kollar, T.; Burchfiel, B.; Tedrake, R.; Sadigh, D.; Levine, S.; Liang, P.; and Finn, C. 2024. OpenVLA: An Open-Source Vision- Language-Act...

  10. [18]

    D.; Liang, Y.; Huang, J.- B.; and Huang, F

    Lee, S.; Jung, Y.; Chun, I.; Lee, Y.-C.; Cai, Z.; Huang, H.; Talreja, A.; Dao, T. D.; Liang, Y.; Huang, J.- B.; and Huang, F. 2025. TraceGen: World Model- ing in 3D Trace Space Enables Learning from Cross- Embodiment Videos. arXiv:2511.21690

  11. [19]

    Li, F.; Song, W.; Zhao, H.; Wang, J.; Ding, P.; Wang, D.; Zeng, L.; and Li, H. 2025. Spatial Forcing: Im- plicit Spatial Representation Alignment for Vision- Language-Action Model. arXiv:2510.12276

  12. [20]

    Lin, M.; Ding, P.; Wang, S.; Zhuang, Z.; Liu, Y.; Tong, X.; Song, W.; Lyu, S.; Huang, S.; and Wang, D. 2026. HiF-VLA: Hindsight, Insight and Foresight through Motion Representation for Vision-Language- Action Models. arXiv:2512.09928

  13. [21]

    Lipman, Y.; Chen, R. T. Q.; Ben-Hamu, H.; Nickel, M.; and Le, M. 2023. Flow Matching for Generative Modeling. InThe Eleventh International Conference on Learning Representations

  14. [22]

    Liu,B.;Zhu,Y.;Gao,C.;Feng,Y.;Liu,Q.;Zhu,Y.;and Stone, P. 2023. LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning. InAdvances in Neural Information Processing Systems 36, 44776– 44791

  15. [23]

    Lu,J.;Xu,J.;Hu,W.;Zhu,R.;Zhao,C.;Yeung,S.-K.; Shan, Y.; and Liu, Y. 2026. Track4World: Feedfor- ward World-Centric Dense 3D Tracking of All Pixels. arXiv:2603.02573

  16. [24]

    F1:A Vision-Language-Action Model Bridging Understand- ing and Generation to Actions

    Lv, Q.; Kong, W.; Li, H.; Zeng, J.; Qiu, Z.; Qu, D.; Song,H.;Chen,Q.;Deng,X.;andPang,J.2025. F1:A Vision-Language-Action Model Bridging Understand- ing and Generation to Actions. arXiv:2509.06951

  17. [25]

    GR00T N1: AnOpen Foundation Model for Generalist Humanoid Robots

    NVIDIA; Bjorck,J.; et al.2025. GR00T N1: AnOpen Foundation Model for Generalist Humanoid Robots. arXiv:2503.14734

  18. [26]

    Open X-Embodiment Collaboration; O’Neill, A.; et al. 2024. Open X-Embodiment: Robotic Learning DatasetsandRT-XModels.In2024IEEEInternational Conference on Robotics and Automation, 6892–6903

  19. [27]

    Pai, J.; Achenbach, L.; Montesinos, V.; Forrai, B.; Mees, O.; and Nava, E. 2025. mimic-video: Video- ActionModelsforGeneralizableRobotControlBeyond VLAs. arXiv:2512.15692

  20. [28]

    Pan, S.; Xu, Y.; Xu, R.; Zhou, Z.; Wu, S.; and Yu, Z. 2025. Self-Correcting Robot Manipulation via Gaussian-Splatted Foresight.Proceedings of the AAAI Conference on Artificial Intelligence, 39(25): 26642– 26650

  21. [29]

    ScalableDiffusionMod- elswithTransformers.InProceedingsoftheIEEE/CVF International Conference on Computer Vision

    Peebles,W.;andXie,S.2023. ScalableDiffusionMod- elswithTransformers.InProceedingsoftheIEEE/CVF International Conference on Computer Vision

  22. [30]

    Pertsch,K.;Stachowicz,K.;Ichter,B.;Driess,D.;Nair, S.; Vuong, Q.; Mees, O.; Finn, C.; and Levine, S

  23. [31]

    GeoPredict:LeveragingPredic- tiveKinematicsand3DGaussianGeometryforPrecise VLA Manipulation

    Qian, J.; Han, B.; Shi, C.; Xiao, L.; Yang, L.; Shi, S.;andJiang,L.2026. GeoPredict:LeveragingPredic- tiveKinematicsand3DGaussianGeometryforPrecise VLA Manipulation. arXiv:2512.16811

  24. [32]

    Qu, D.; Song, H.; Chen, Q.; Yao, Y.; Ye, X.; Gu, J.; Wang, Z.; Ding, Y.; Zhao, B.; Wang, D.; and Li, X

  25. [33]

    Su, Y.; Chen, S.; Shi, H.; Liu, M.; Zhang, Z.; Huang, N.; Zhong, W.; Zhu, Z.; Liu, Y.; and Liu, X. 2026. World Guidance: World Modeling in Condition Space for Action Generation. arXiv:2602.22010

  26. [34]

    InRobotics: Sci- ence and Systems XXI

    SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Models. InRobotics: Sci- ence and Systems XXI

  27. [35]

    Wang, X.; Wang, C.; Xu, Y.; Ye, M.; Zhang, F.; Tian, J.; Zhan, X.; Zhu, L.; Lu, C.; and Yang, L

  28. [36]

    Sun,J.;Zhang,W.;Qi,Z.;Ren,S.;Liu,Z.;Zhu,H.;Sun, G.; Jin, X.; and Chen, Z. 2026. VLA-JEPA: Enhanc- ing Vision-Language-Action Model with Latent World Model. arXiv:2602.10098. 9

  29. [37]

    Zheng, J.; Li, J.; Wang, Z.; Liu, D.; Kang, X.; Feng, Y.; Zheng, Y.; Zou, J.; Chen, Y.; Zeng, J.; Zhang, Y.-Q.; Pang, J.; Liu, J.; Wang, T.; and Zhan, X

  30. [38]

    Zheng, R.; Liang, Y.; Huang, S.; Gao, J.; Daumé III, H.; Kolobov, A.; Huang, F.; and Yang, J. 2024. TraceVLA: Visual Trace Prompting Enhances Spatial- Temporal Awareness for Generalist Robotic Policies. arXiv:2412.10345

  31. [39]

    Zhen, H.; Qiu, X.; Chen, P.; Yang, J.; Yan, X.; Du, Y.; Hong, Y.; and Gan, C. 2024. 3D-VLA: A 3D Vision-Language-Action Generative World Model. In Proceedings of the 41st International Conference on Machine Learning

  32. [40]

    Zhong, Z.; Yan, H.; Li, J.; Liu, X.; Gong, X.; Zhang, T.; Song, W.; Chen, J.; Zheng, X.; Wang, H.; and Li, H. 2025. FlowVLA: Visual Chain of Thought-based Motion Reasoning for Vision-Language-Action Mod- els. arXiv:2508.18269. 10

  33. [41]

    arXiv:2510.10274

    X-VLA:Soft-PromptedTransformerasScalable Cross-Embodiment Vision-Language-Action Model. arXiv:2510.10274

  34. [43]

    Zhi, H.; Chen, P.; Zhou, S.; Dong, Y.; Wu, Q.; Han, L.; and Tan, M. 2025. 3DFlowAction: Learning Cross-EmbodimentManipulationfrom3DFlowWorld Model. arXiv:2506.06199

  35. [2025]

    arXiv:2501.09747

    FAST: Efficient Action Tokenization for Vision- Language-Action Models. arXiv:2501.09747

  36. [2026]

    arXiv:2603.25399

    LaMP: Learning Vision-Language-Action Pol- icy with 3D Scene Flow as Latent Motion Prior. arXiv:2603.25399

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.