Pith. sign in

REVIEW 3 major objections 4 minor 18 references

A world-action model's RGB video branch can be redirected from photorealistic future reconstruction toward tracked gripper, object, and contact motion, improving out-of-distribution robot manipulation without adding modalities or online inp

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

2026-08-01 01:06 UTC pith:7MOWNIVT

load-bearing objection A clean, well-controlled method paper whose central mechanism would be more convincing with tracker validation and variance reporting; worth peer review. the 3 major comments →

arxiv 2607.25918 v1 pith:7MOWNIVT submitted 2026-07-28 cs.RO

DC-WAM: Dynamic-Centric Visual Supervision and Reasoning for World-Action Models

Haoyuan Ji , Lingxiang Fan , Shang Su , Yinqiao Lu , Mengkai Shi , Jun Gao , Shuo Feng This is my paper
classification cs.RO
keywords world-action modelsrobot manipulationvisual foresightflow matchingpoint trajectoriesout-of-distribution robustnessattention biasdynamic-centric supervision
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.

This paper argues that the payoff of giving a robot policy future-video prediction is not the rendered future itself but the control-relevant visual representations the prediction forces the policy to learn. It proposes DC-WAM, which keeps the RGB video branch unchanged at deployment but changes what it is trained to do: instead of reconstructing appearance, the branch is pushed toward interaction-induced dynamics—gripper motion, displaced objects, contact changes—through temporal-difference flow matching and a trajectory-weighted loss. A learned attention bias (DynaRoute) then points the visual encoder at the most dynamic future tokens. Trained only on clean demonstrations, DC-WAM reaches 98.1% on LIBERO and 60.9% on the out-of-distribution LIBERO-Plus, beating the matched action-conditioned baseline by 1.4 and 7.1 points and shrinking the average ID-OOD drop from 42.9 to 37.2 points; real-world bimanual tasks under lighting and background changes follow the same pattern. If this is right, appearance-heavy future prediction is largely wasted capacity for control, and steering video supervision toward tracked motion is a cheap route to OOD robustness.

Core claim

The visual branch of a world-action model should not be treated as a prediction target to be rendered believably; it is a training-time curriculum for action representations. DC-WAM establishes this by redirecting an existing RGB-based WAM from appearance-dominated reconstruction toward interaction-induced visual dynamics without introducing additional modality-specific predictions or online inputs. The mechanism is a tracker-derived dynamic map: point trajectories are computed offline on clean videos, thresholded to keep only moving points, rasterized onto the visual-token grid, and normalized per episode; this map defines both a reweighted flow-matching loss (TrackFM) that concentrates sup

What carries the argument

The central object is the tracker-derived dynamic map, an episode-normalized token-level saliency map built offline from point trajectories estimated by an off-the-shelf tracker on the clean training videos. Candidate points are sampled uniformly or from foreground regions; frame-wise motion magnitudes are thresholded to isolate gripper, manipulated-object, and contact motion; Gaussian kernels rasterize the moving points onto the VAE token grid; and episode-level normalization preserves both spatial and temporal saliency. The same map plays two roles: it reweights the visual flow-matching error at sparse interaction locations (TrackFM) and supervises DynaRoute's token-wise relevance predicti

Load-bearing premise

Everything rests on offline point tracking of the clean training videos: the dynamic map that reweights supervision and drives attention is trustworthy only when the point tracker reliably localizes gripper, object, and contact motion despite occlusion and small-object manipulation; if the tracker errs in ways that correlate with task difficulty or appearance, the whole dynamic-centric signal misguides the visual branch.

What would settle it

Take the clean LIBERO training set, build the dynamic map, then train DC-WAM with the map spatially shifted by, say, half the token-grid size while keeping everything else identical. If LIBERO-Plus success stays near 60.9% instead of falling toward the baseline 53.8%, the trajectory-guided reweighting and DynaRoute routing are not actually using the tracked interaction locations and the central claim fails.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the central claim holds, future-video fidelity is not the right objective for WAM visual branches; control-relevant weighting is what matters.
  • Trained only on clean demonstrations, DC-WAM raises LIBERO-Plus success from 53.8% to 60.9% and cuts the average ID-OOD drop from 42.9 to 37.2 points.
  • The same scheme transfers to real-world bimanual manipulation, improving success under lighting shifts and background distractors with smaller clean-to-OOD drops than the baseline.
  • Since dynamic maps are built offline, the robot needs no tracker, segmenter, or structured future predictions at execution; inference remains action-only with a routed video cache.
  • The dense temporal-difference and sparse trajectory-weighted losses are complementary: together they beat either alone, and DynaRoute's gain requires spatially aligned relevance rather than arbitrary or action-query routing.

Where Pith is reading between the lines

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

  • If the representation story is right, one could slim the video branch further—or even distill the dynamic focus into a single forward pass—since the rendered future is incidental; a testable prediction is that a frozen DC-WAM video encoder transfers better to a much smaller action policy than an appearance-trained encoder.
  • The principle generalizes: any auxiliary prediction target used to regularize a policy should be weighted by its action-relevance. Applying the same reweighting to depth, tactile, or semantic predictions could yield analogous robustness without adding deployment sensors.
  • The remaining risk is downstream of the tracker: on tasks where point trajectories drift under occlusion, fast motion, or small-object contact, the dynamic map may mislocalize the interaction and the gains could invert; evaluating DC-WAM under deliberately corrupted or withheld trajectory maps would map this boundary.
  • An implicit claim is that appearance shifts like lighting do not alter interaction trajectories; this suggests a cheap OOD test: perturb the appearance of the training video only, freeze trajectories, and verify that policy success is preserved—if it drops, the visual branch is still absorbing appearance signals.

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. DC-WAM is a training-time modification of an RGB video-prediction branch inside a World-Action Model (WAM). The authors propose (i) dynamic-centric visual supervision consisting of a temporal-difference flow-matching loss (Eq. 22) and a tracker-guided reweighted flow-matching loss (Eq. 24), where the weighting map m* is built offline from CoTracker3 point trajectories (Eqs. 2–7); and (ii) DynaRoute, a lightweight module that predicts token-wise dynamic relevance and injects it as an attention bias into the visual branch (Eqs. 10–14). At deployment, the video branch is executed once to produce a routed key-value cache and is then disabled, following Fast-WAM-style action-only inference. Experiments on LIBERO/LIBERO-Plus and three real-world bimanual tasks report consistent improvements over matched FastWAM and FastWAM-AC baselines, especially under lighting, background, and language perturbations. The central claim is that an existing RGB-based WAM can be redirected from appearance-dominated reconstruction toward interaction-induced dynamics without additional modality-specific predictions or online inputs at deployment.

Significance. If the central claim holds, DC-WAM makes a useful contribution to the WAM literature: it shows that the RGB video branch can be steered toward control-relevant dynamics using only offline trajectory-derived targets, while preserving efficient action-only inference. The paper has several strengths: matched baselines under the same backbone, coherent ablations separating dense vs. sparse supervision and routing variants, and real-world evaluation under unseen perturbations. The claims are specific and falsifiable. However, the load-bearing mechanism — that the tracker-derived dynamic maps actually localize interaction-induced motion — is not validated, and the reported gains lack statistical confidence intervals. Until these are addressed, the magnitude and mechanism of the improvements remain uncertain.

major comments (3)
  1. [Method, Eqs. (2)–(7), (24)–(25)] The entire method is downstream of the tracker-derived dynamic map m*. Both TrackFM (Eq. 24) and the DynaRoute target (Eq. 25) are computed from CoTracker3 trajectories on clean training videos. The paper explicitly conditions its robustness claim on the assertion 'provided that the underlying point trajectories remain stable,' but this condition is never tested. In contact-heavy tasks such as opening a basket or stacking plates, point trackers frequently lose tracks at occlusion and contact, and the δ_mot threshold in Eq. (4) can select robot-arm motion rather than object/contact motion. No tracker accuracy analysis, no oracle-map comparison, and no failure-case study is reported. If m* is noisy or task-correlated, the empirical gains could arise from generic regularization rather than the claimed dynamic-centric mechanism. Please add (a) quantitative validation of CoTracker3 trajectori
  2. [Tables 1 and 2] No standard deviations, confidence intervals, or per-seed results are reported for any experiment. With 50 rollouts per LIBERO task, the binomial standard error is roughly 3–5 percentage points for the observed success rates, so several of the headline deltas (e.g., +1.4 on LIBERO, +8 and +9 on clean real-world tasks in Table 2) are within about 1–1.5 standard errors. LIBERO-Plus deltas of 7–9 points are more convincing, but still need variance estimates. Real-world results with 100 trials per condition also lack CIs. Please report per-seed means with standard deviations or at least bootstrap 95% CIs, and specify the number of seeds. Without this, the claim of 'consistently improves' is not statistically supported.
  3. [Eq. (17) and 'Action-only inference with routed video cache'] DynaRoute is trained with inputs at all diffusion timesteps, including lower-noise regimes where future-visual tokens contain partially preserved structure, but at deployment it is evaluated once on pure noise future slots at τ_init = 1. The paper asserts this 'closely matches the high-noise regime used during training' without quantitative support. If DynaRoute's relevance predictions are inaccurate on pure-noise inputs, the cached bias could be near-random. Please ablate τ_init (e.g., compare τ_init = 0.2, 0.5, 1.0), or train DynaRoute only on high-noise timesteps and compare, and report the routing bias accuracy on pure-noise inputs. This is directly load-bearing for the deployment-time behavior.
minor comments (4)
  1. [Fig. 2] Figure 2 contains garbled notation (e.g., '??', 'J`') and should be cleaned before publication. The pipeline would also benefit from a clear label indicating where m* is computed offline vs. used online.
  2. [Eqs. (12)–(14)] DynaRoute outputs relevance over all Sv visual tokens (including observation tokens), but Eq. (12) describes the bias only for future visual tokens. Eq. (13) centers over all Sv tokens. Please clarify whether observation tokens receive the bias and how the centering affects the future-token bias values. This affects interpretability of the attention-bias plots.
  3. [Real-world experimental setup] The real-world section is under-specified. It reports 100 demonstrations per task and 100 trials per task–condition pair, but omits episode horizons, action dimensionality, control frequency, camera setup, and whether the same demonstration data are used for all methods. Please provide these details and report per-condition rollouts.
  4. [Reproducibility] No code or checkpoint release is mentioned. Given the dependence on CoTracker3, SAM, Wan2.2, and LIBERO-Plus, releasing the dynamic-map construction pipeline and DynaRoute implementation would substantially aid reproducibility.

Circularity Check

0 steps flagged

No circularity: tracker-derived dynamic maps are offline external targets; policy success is measured on held-out rollouts, and the self-citations are not load-bearing.

full rationale

The derivation chain is self-contained with respect to its targets. The dynamic map m* is constructed offline from CoTracker3 point trajectories on ground-truth training videos (Eqs. 2–7) and is used as a fixed supervision target: Eq. 24 reweights the flow-matching error by m*, and Eq. 25 trains DynaRoute to predict m*. Neither operation defines its output as its input; g is a learned network output, not a copy of m*. The central empirical claims (Tables 1–2, Fig. 1) are evaluated on held-out LIBERO/LIBERO-Plus rollouts and real-world trials, not on training objectives. The only coauthored references (Gao et al. 2026; Ye et al. 2026b) are background or attribution for a loss whose equation is stated in the paper; they do not supply a load-bearing premise. The paper's own caveat that robustness holds 'provided that the underlying point trajectories remain stable' is a correctness condition about the external tracker, not a circular reduction: if CoTracker3 fails, the supervision targets are misinformed, but that is a data-quality issue rather than a logical equivalence. No self-definition, fitted-input-called-prediction, imported uniqueness, or ansatz-via-citation step was found.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

No new physical entities or conserved quantities are introduced; DynaRoute and the dynamic maps are model components or derived training targets, not postulated entities. The central claim rests on standard flow-matching math plus several domain assumptions about tracker quality, segmentation quality, spatial locality of VAE latents, and dynamics invariance under appearance shifts. The main hand-tuned quantities are the kernel parameters and routing/loss weights.

free parameters (6)
  • δmot (motion threshold for dynamic point selection)
    Eq. (4): points with d_t,n > δmot are called dynamic; value not reported, and it controls which regions get supervision weight.
  • λ (kernel bandwidth factor) = 0.25
    Eq. (5): set by hand for Gaussian rasterization of tracked points onto VAE tokens.
  • σp (kernel scale parameter) = 1.25
    Eq. (5) and text: σx = (W/Wz)σp, σy = (H/Hz)σp; chosen by hand.
  • α (DynaRoute routing strength)
    Eq. (12): converts predicted relevance into log-space attention bias; value not reported; controls how strongly low-relevance tokens are suppressed.
  • λ_Dice (Dice loss weight)
    Eq. (25): weight combining BCE and soft Dice for L_Route; value not reported.
  • loss weights of L_A_FM, L_V_TD, L_V_TrackFM, L_Route
    Eq. (18): total objective is an unweighted-looking sum but no coefficients are stated; if tuned, they affect balance of action vs visual supervision.
axioms (5)
  • domain assumption CoTracker3 point trajectories in clean training videos correctly localize gripper, object, and contact motion.
    Eqs. (2)-(7); all dynamic maps, TrackFM weights, and DynaRoute targets are derived from these trajectories; no tracker failure analysis for occlusion/contact is provided.
  • domain assumption SAM foreground regions improve dynamic-point coverage over uniform sampling.
    Method section: 'candidate points either uniformly over the image or within foreground regions produced by SAM'; if SAM mis-segments, dynamic maps can miss interaction regions.
  • standard math The flow-matching linear interpolation xτ=(1−τ)x+τε with target u=ε−x is a valid training objective for the visual branch.
    Eqs. (19)-(20); standard flow-matching from prior literature, not derived here.
  • domain assumption Appearance-level OOD shifts do not alter the underlying manipulation dynamics.
    Central OOD robustness claim (intro and analysis); if lighting/background changes affect point-trajectory stability in training or dynamics at test, dynamic-centric reweighting may not transfer.
  • domain assumption The Wan2.2 VAE and DiT token grids preserve spatial locality well enough for token-grid dynamic maps to be meaningful.
    Eqs. (5)-(8); the Gaussian rasterization and downsampling assume the latent token grid aligns spatially with image coordinates.

pith-pipeline@v1.3.0-alltime-deepseek · 11083 in / 12600 out tokens · 107006 ms · 2026-08-01T01:06:24.449889+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of DC-WAM: Dynamic-Centric Visual Supervision and Reasoning for World-Action Models." pith.science (2026). https://pith.science/paper/7MOWNIVT

@misc{pith2026260725918,
  author       = {Pith},
  title        = {Pith review of: DC-WAM: Dynamic-Centric Visual Supervision and Reasoning for World-Action Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7MOWNIVT}},
  note         = {Machine review of arXiv:2607.25918}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

World-Action Models (WAMs) augment robot policies with future visual prediction, but it remains unclear what the visual modality should learn for control. While photorealistic future prediction provides dense supervision, it also incurs substantial computation and can allocate capacity to texture, illumination, and background variations that are only weakly related to action selection. Recent efficient WAM variants suggest that the main benefit of the video branch may not lie in the rendered future itself, but in the control-relevant visual representations induced during training. In this work, we revisit future video prediction from a dynamic-centric perspective and ask whether an existing RGB-based WAM can be redirected from appearance-dominated reconstruction toward interaction-induced visual dynamics without introducing additional modality-specific predictions or online inputs at deployment. We propose DC-WAM, a dynamic-centric WAM framework that redistributes supervision and computation in the RGB video branch. At the supervision level, DC-WAM combines temporal-difference flow matching with trajectory-guided weighting, emphasizing dense temporal changes and localized regions where the gripper, manipulated objects, and contact areas move. At the reasoning level, DynaRoute predicts token-wise dynamic relevance and converts it into an attention bias, guiding the model toward control-relevant future tokens. Experiments in simulation and on real-world manipulation tasks show that DC-WAM consistently improves policy performance, especially under out-of-distribution perturbations in lighting, object appearance, and background texture.

Figures

Figures reproduced from arXiv: 2607.25918 by Haoyuan Ji, Jun Gao, Lingxiang Fan, Mengkai Shi, Shang Su, Shuo Feng, Yinqiao Lu.

Figure 1
Figure 1. Figure 1: Effect of dynamic-centric supervision and routing. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of DC-WAM. DC-WAM redirects the RGB video branch toward interaction-induced dynamics through [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Attention responses under increasing visual perturbations. As perturbation severity increases, FastWAM-AC attention [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: DynaRoute during training and inference. During [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Real-world task-by-condition evaluation ma [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

18 extracted references · 18 linked inside Pith

  1. [1]

    X.; Tanner, J.; Vuong, Q.; Walling,A.;Wang,H.;andZhilinsky,U.2024.π 0:AVision- Language-Action Flow Model for General Robot Control

    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. X.; Tanner, J.; Vuong, Q.; Walling,A.;Wang,H.;andZhilinsky,U.2024.π 0:AVision- Language-Action Flow Model for General Robot Control. arXiv preprint arXiv:2410...

    Pith/arXiv arXiv 2024

  2. [3]

    Gao, S.; Liang, W.; Zheng, K.; Malik, A.; Ye, S.; Yu, S.; Tseng, W.-C.; Dong, Y.; Mo, K.; Lin, C.-H.; Ma, Q.; Nah, S.; Magne, L.; Xiang, J.; Xie, Y.; Zheng, R.; Niu, D.; Tan, Y

    LIBERO-Plus: In-Depth Robustness Analysis of Vision- Language-ActionModels.arXiv preprint arXiv:2510.13626. Gao, S.; Liang, W.; Zheng, K.; Malik, A.; Ye, S.; Yu, S.; Tseng, W.-C.; Dong, Y.; Mo, K.; Lin, C.-H.; Ma, Q.; Nah, S.; Magne, L.; Xiang, J.; Xie, Y.; Zheng, R.; Niu, D.; Tan, Y. L.; Zentner, K. R.; Kurian, G.; Indupuru, S.; Jannaty, P.; Gu, J.; Zhan...

    Pith/arXiv arXiv

  3. [6]

    arXiv preprint arXiv:2605.23856

    Point Tracking Improves World Action Models. arXiv preprint arXiv:2605.23856. Hu, Y.; Guo, Y.; Wang, P.; Chen, X.; Wang, Y.-J.; Zhang, J.; Sreenath, K.; Lu, C.; and Chen, J

    Pith/arXiv arXiv

  4. [8]

    CoTracker3: Simpler and Bet- ter Point Tracking by Pseudo-Labelling Real Videos.arXiv preprint arXiv:2410.11831. Kim, M. J.; Gao, Y.; Lin, T.-Y.; Lin, Y.-C.; Ge, Y.; Lam, G.; Liang, P.; Song, S.; Liu, M.-Y.; Finn, C.; and Gu, J

    Pith/arXiv arXiv

  5. [9]

    arXiv:2601.16163

    Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning. arXiv:2601.16163. Kim, M. J.; Pertsch, K.; Karamcheti, S.; Xiao, T.; Balakr- ishna,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

    Pith/arXiv arXiv

  6. [10]

    Kirillov, A.; Mintun, E.; Ravi, N.; Mao, H.; Rolland, C.; Gustafson, L.; Xiao, T.; Whitehead, S.; Berg, A

    OpenVLA: An Open-Source Vision-Language-Action Model.arXiv preprint arXiv:2406.09246. Kirillov, A.; Mintun, E.; Ravi, N.; Mao, H.; Rolland, C.; Gustafson, L.; Xiao, T.; Whitehead, S.; Berg, A. C.; Lo, W.- Y.; Dollár, P.; and Girshick, R

    Pith/arXiv arXiv

  7. [11]

    arXiv preprint arXiv:2304.02643

    Segment Anything. arXiv preprint arXiv:2304.02643. Li, J.; Guo, T.; Ye, Y.; Zhang, R.; Chi, X.; Sun, Q.; Li, Y.; Lou, Y.; Huang, Y.; Lu, Z.; Guo, M.; and Zhang, S. 2026a. Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination.arXiv preprint arXiv:2606.10040. Li, L.; Zhang, Q.; Luo, Y.; Yang, S.; Wang, R.; Han, F.; Yu, M.; Gao, Z...

    Pith/arXiv arXiv 2025

  8. [12]

    Liu, Y.; Sun, P.; Li, S.; Xie, Y.; Zhang, L.; Chao, X.; Dong, S.; Chen, F.; Zhang, X.-P.; and Ding, W

    LIBERO: Benchmarking Knowl- edge Transfer for Lifelong Robot Learning.arXiv preprint arXiv:2306.03310. Liu, Y.; Sun, P.; Li, S.; Xie, Y.; Zhang, L.; Chao, X.; Dong, S.; Chen, F.; Zhang, X.-P.; and Ding, W

    Pith/arXiv arXiv

  9. [13]

    Lou,Y.;Chi,X.;Zhang,X.;Qian,Z.;Li,C.;Zhang,R.;Lyu, Y.; Song, G.; Fu, C.; Xu, H.; Wang, P.; and Zhang, S

    OA-WAM: Object-Addressable World Action Model for Robust Robot Manipulation.arXiv preprint arXiv:2605.06481. Lou,Y.;Chi,X.;Zhang,X.;Qian,Z.;Li,C.;Zhang,R.;Lyu, Y.; Song, G.; Fu, C.; Xu, H.; Wang, P.; and Zhang, S

    Pith/arXiv arXiv

  10. [14]

    Press,O.;Smith,N.A.;andLewis,M.2022.TrainShort,Test Long: Attention with Linear Biases Enables Input Length Extrapolation

    Mask World Model: Predicting What Matters for Robust Robot Policy Learning.arXiv preprint arXiv:2604.19683. Press,O.;Smith,N.A.;andLewis,M.2022.TrainShort,Test Long: Attention with Linear Biases Enables Input Length Extrapolation. arXiv:2108.12409. Ranasinghe,K.;Zhou,H.;Fang,Y.;Yang,L.;Xue,L.;Xu,R.; Xiong,C.;Savarese,S.;Ryoo,M.S.;andNiebles,J.C.2026. Futu...

    Pith/arXiv arXiv 2022

  11. [15]

    arXiv:2503.20314

    Wan: Open and Advanced Large-Scale Video Generative Models. arXiv:2503.20314. Ye, A.; Wang, B.; Ni, C.; Huang, G.; Zhao, G.; Li, H.; Li, H.; Li, J.; Lv, J.; Liu, J.; Cao, M.; Li, P.; Deng, Q.; Mei, W.; Wang, X.; Chen, X.; Zhou, X.; Wang, Y.; Chang, Y.; Li, Y.; Zhou, Y.; Ye, Y.; Liu, Z.; and Zhu, Z. 2026a. GigaWorld- Policy: An Efficient Action-Centered Wo...

    Pith/arXiv arXiv

  12. [16]

    Yuan, T.; Dong, Z.; Liu, Y.; and Zhao, H

    MaskWAM: Unifying Mask Prompting and Prediction for World-Action Models.arXiv preprint arXiv:2606.13515. Yuan, T.; Dong, Z.; Liu, Y.; and Zhao, H

    Pith/arXiv arXiv

  13. [17]

    Zhang, J.; Zhu, J.; Su, T.; Ma, C.; Huang, Z.; Xu, Y.; and Wang, H

    Fast-WAM: Do World Action Models Need Test-Time Future Imagina- tion?arXiv preprint arXiv:2603.16666. Zhang, J.; Zhu, J.; Su, T.; Ma, C.; Huang, Z.; Xu, Y.; and Wang, H. 2026a. Learning 4D Geometric Priors for Inference-Efficient World Action Models.arXiv preprint arXiv:2607.05468. Zhang, P.; Deng, Y.; Sun, S.; Ma, J.; Wang, D.; Du, J.; Pan, Z.; Huang, Y....

    Pith/arXiv arXiv

  14. [18]

    Zhu,C.;Yu,R.;Feng,S.;Burchfiel,B.;Shah,P.;andGupta, A.2025

    TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Aware- ness for Generalist Robotic Policies.arXiv preprint arXiv:2412.10345. Zhu,C.;Yu,R.;Feng,S.;Burchfiel,B.;Shah,P.;andGupta, A.2025. UnifiedWorldModels:CouplingVideoandAction Diffusion for Pretraining on Large Robotic Datasets.arXiv preprint arXiv:2504.02792

    Pith/arXiv arXiv 2025

  15. [2023]

    Guan,J.;Zhao,W.;Pei,Y.;Chen,Z.;Solin,A.;andKannala, J

    RT-Trajectory: Robotic Task Gener- alization via Hindsight Trajectory Sketches.arXiv preprint arXiv:2311.01977. Guan,J.;Zhao,W.;Pei,Y.;Chen,Z.;Solin,A.;andKannala, J

    Pith/arXiv arXiv

  16. [2024]

    Karaev, N.; Makarov, I.; Wang, J.; Neverova, N.; Vedaldi, A.; and Rupprecht, C

    Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations.arXiv preprint arXiv:2412.14803. Karaev, N.; Makarov, I.; Wang, J.; Neverova, N.; Vedaldi, A.; and Rupprecht, C

    Pith/arXiv arXiv

  17. [2025]

    Fei,S.;Wang,S.;Shi,J.;Dai,Z.;Cai,J.;Qian,P.;Ji,L.;He, X.; Zhang, S.; Fei, Z.; Fu, J.; Gong, J.; and Qiu, X

    AimBot: A Simple Auxiliary Visual Cue to Enhance Spatial Awareness of Visuomotor Policies.arXiv preprint arXiv:2508.08113. Fei,S.;Wang,S.;Shi,J.;Dai,Z.;Cai,J.;Qian,P.;Ji,L.;He, X.; Zhang, S.; Fei, Z.; Fu, J.; Gong, J.; and Qiu, X

    Pith/arXiv arXiv

  18. [2026]

    Gu, J.; Kirmani, S.; Wohlhart, P.; Lu, Y.; Arenas, M

    DreamDojo: A Generalist RobotWorldModelfromLarge-ScaleHumanVideos.arXiv preprint arXiv:2602.06949. Gu, J.; Kirmani, S.; Wohlhart, P.; Lu, Y.; Arenas, M. G.; Rao, K.; Yu, W.; Fu, C.; Gopalakrishnan, K.; Xu, Z.; Sun- daresan, P.; Xu, P.; Su, H.; Hausman, K.; Finn, C.; Vuong, Q.; and Xiao, T

    Pith/arXiv arXiv