Pith.

Robotics · arXiv:2610.30991 v1 · submitted 22 Oct 2026 · 51 references

UniFlowWAM: A Unified World-Action Model with Visual Semantics and 4D Geometry

Jialin Tian, Zhuofan Li, Xinkai Wang, Bokai Lin, Chenyi Wang, Xinyu Zhan, Cewu Lu, Lixin Yang

The paper says

“Robot demonstrations record what the robot saw and what it did. They do not specify how the 3D world changed in between. Scene-level 3D flow describes it by tracking 3D scene points, providing the world representation with 4D geometry.”

more of the abstract

Prior work has used flow to improve robot policies. Flow served as an intermediate prediction that a separate model converts into actions, as a prior injected at inference time, or as an extra output learned together with the action. In all of these designs, flow stays outside the policy, as a separate module or as an extra prediction head. We introduce UniFlowWAM, a Unified World-Action Model that makes 3D scene flow a third modality on par with vision and action. Its world representation therefore covers both visual semantics, through future visual features, and 4D geometry, through 3D flow, and flow receives its own supervision. UniFlowWAM is a diffusion transformer with one expert per modality, and the three experts exchange information through joint attention. We train UniFlowWAM under a ten-objective role grid, where each objective assigns every modality exactly one of three roles, namely target, condition, or absent. The same model thus serves as a policy, as forward and inverse dynamics, and as a predictor of future visual features or 3D flow. We evaluate UniFlowWAM on RoboCasa-GR1, LIBERO, and LIBERO-Plus in simulation, and on four real-world bimanual manipulation tasks. On RoboCasa-GR1, UniFlowWAM raises the success rate of its base model without flow from 52.1% to 56.4%. It preserves in-distribution success on LIBERO at 98.1% and reaches 87.7% zero-shot on LIBERO-Plus. On the real robot, it achieves 87.0% average success, compared with 68.5% for the strongest baseline.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 23 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A solid empirical contribution that makes 3D flow a co-equal world-model modality; the 25-trial real-robot ablation needs confidence intervals and a cleaner schedule comparison before it can carry the load. This line is wrong

4 major objections · 3 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 51 references say.

The paper says, at Sec. 4.4, Table 4

The review did not quote this passage.

Pith says: Objection 1 of 4 · major

The ablation that is load-bearing for the paper's central claim is confounded by training schedule. The text states that the reduced objective sets level off within 120k steps while the full ten-objective grid is still improving at 120k steps, and that the full grid is therefore extended to 200k steps; however, Table 4 then reports the reduced sets at 120k and the full-grid row from the separate 50-trial main run, not a schedule-matched comparison.

more of objection 1

The sentence "at equal training steps the full grid clearly outperforms policy-only training" is asserted but is not what Table 4 displays. Please report all four objective variants at matched steps (for example, 120k and 200k) or otherwise separate the effect of the objective set from the effect of the longer optimization budget.

The paper says, at Sec. 4.4, Table 4

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

The real-robot objective ablation is underpowered as reported. The three ablation rows are measured over 25 trials per task with no confidence intervals, and the per-orientation subgroup used to support the failure-distribution argument contains 5 trials per variant. In the 90-degree chili group, 1/5 versus 4/5 successes has a two-sided Fisher exact test p-value of roughly 0.2, so the text's …

more of objection 2

… statement that a gap of 20% versus 80% is "large enough that we read it as a real difference rather than as sampling noise" is not a statistical justification. Please provide confidence intervals, exact trial counts per condition, or additional trials for the decisive comparisons, and apply a formal test to the subgroup claim.

The paper says, at Sec. 4.2.2, Table 3

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

The LIBERO-Plus headline of 87.65% is reported from the best evaluated checkpoint rather than from a prespecified final checkpoint. With the official zero-shot protocol, reporting the best checkpoint can inflate the result relative to baselines whose numbers may come from a fixed training schedule. Please report the checkpoint-selection rule, the last-checkpoint or average-over-checkpoints performance, and the variance across seeds, so that the robustness claim in Contribution (iii) is comparable to the baselines in Table 3.

The paper says, at Sec. 3.2 and Sec. 5

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.65):

“These labels can contain temporal inconsistencies, such as an isolated reversal of flow direction within a continuous trajectory.”

Pith says: Objection 4 of 4 · major

The mechanism attributed to the flow modality is not directly verified. The paper acknowledges that Track4World pseudo-labels can contain temporal inconsistencies and that the 4D VAE compresses foreground-only flow after SAM segmentation, but it provides no quantitative check of label fidelity on the robot demonstration data, no reconstruction or tracking-quality numbers, and no experiment that corrupts or ablates the flow labels.

more of objection 4

Because flow is never consumed at test time, the observed gains could in principle come from the auxiliary-task regularization of predicting any additional target. Please add a label-quality evaluation, a comparison of predicted flow latents against held-out pseudo-labels, or a control experiment with corrupted flow targets to support the claim that 4D geometry, rather than generic multi-task regularization, drives the improvement.

3 minor objections

The paper says, at Table 2

Pith says: minor objection 1 of 3

The upper block of Table 2 mixes UniFlowWAM's best result over evaluation seeds (58.9%, marked with a dagger) with published numbers from other methods whose selection rule is not stated; the repeated-evaluation block below is clear, but the comparison block should specify how each baseline number was obtained.

The paper says, at Abstract and Sec. 1

Pith says: minor objection 2 of 3

The abstract says flow is "on par" with vision and action, while Sec. 5 acknowledges that the flow expert has not been pretrained at the same scale and uses roughly 1,200 hours of demonstrations versus the much larger LDA pretraining corpus; the wording could be tempered to match the stated training asymmetry.

show 1 more

The paper says, at Sec. 4.4

Pith says: minor objection 3 of 3

The sentence "All later experiments, including every UniFlowWAM result in this paper, therefore use no gate" (Sec. 4.5) is clear, but the gating experiment is described as not a controlled ablation; consider labeling it explicitly as a diagnostic in the caption of the corresponding paragraph.

What would settle it

Re-run the three Table 4 ablation rows (policy-only, no-flow, flow-as-condition, full grid) on the ALOHA platform with at least 50 trials per row per task and report per-cell Wilson 95% confidence intervals; minimally, report per-orientation and per-edge subgroup counts with a Fisher exact test for the full-grid versus policy-only comparison. Additionally, train the full-grid variant on the same 120k schedule used for the reduced sets and compare checkpoints at matched steps; if the 12 to 18 point gap collapses at matched training budget, the claim that the flow objectives add value is not supported by the current protocol.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 23 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 17 s · 17,136 tokens read, 2,031 written
  3. Editor wrote the verdict above 32 s · 18,217 tokens read, 4,209 written
  4. Referee major revision: 4 major objections, 3 minor 31 s · 17,500 tokens read, 3,795 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Bottom line: this paper deserves a serious referee, and if the real-robot ablation were cleaned up it could be a straightforward accept. The new idea is real: flow becomes a co-equal modality inside a diffusion policy, organized by a ten-objective role grid that makes flow both a prediction target and a conditioning signal, while deployment needs no flow input. That is genuinely different from prior flow-as-module, flow-as-prior, or flow-as-extra-head designs, and the paper makes the design concrete.

The evaluation is mostly honest and careful. RoboCasa is reported over 11 shared seeds with standard deviations; the gain from 52.11 to 56.38 is modest but consistent and concentrates on the weakest tasks. The LIBERO-Plus result (87.7% zero-shot, 96.9% under camera perturbation versus 76.6% for the best prior) is the strongest single number. The real-robot protocol fixes starting conditions so all policies see the same inputs, and the objective ladder is a sensible ablation. The authors also report discarded designs and state their own limitations; the flow-noise and limited-pretraining caveats are already on the page.

Soft spots, in decreasing order. First, the load-bearing real-robot ablation in Table 4 rests on 25-trial rows and mixes them with the 50-trial main row. The 12 to 18 point gaps are suggestive, but at n=25 the reported precision is thin. The subgroup argument in Sec. 4.4 is weaker than the text implies: 1/5 versus 4/5 success is roughly p=0.2, so that particular failure-distribution story does not add much. The authors should report exact binomial confidence intervals or run more trials. The schedule issue is also muddier than the prose suggests: the full grid alone is extended to 200k in the simulation section while reduced sets stop at 120k. If the obj-10 real-robot row is the 200k run, the comparison is not quite as clean as the table appears to make it. Second, Table 2 reports the best-over-seeds number (58.9) alongside published single numbers and then gives the repeated-seed mean plus standard deviation (56.38); the mixed conventions make the headline hard to read. LIBERO-Plus is reported at the best evaluated checkpoint, which should be labeled clearly. Third, the flow pseudo-labels are acknowledged as noisy. That is fine as a limitation; the concern that the model could be learning tracker errors rather than true 4D motion is real but not demonstrated.

The central claim, that flow's own supervision and conditioning improve the policy, holds in direction. RoboCasa and LIBERO-Plus support it independently of the real-robot run, and the weaknesses are addressable reporting and statistical issues rather than a broken argument. For anyone working on world-action models or flow-based manipulation, this is worth reading and discussing. I would send it to peer review, ask for a cleaned-up ablation table with confidence intervals and explicit training schedules, and expect the revision to be a solid paper.

The strongest objection

The central claim (Sec. 1 Contribution iii; Sec. 4.2) is that 3D flow, as a co-equal modality with its own supervision, improves the action policy. The decisive evidence is Table 4 on the real robot: full-grid post-training beats policy-only by 12 to 18 points on chili transfer, pen manipulation, and towel folding. The paper's own text acknowledges the sampling-noise concern (Sec. 4.4, 'a gap of 20% against 80% is large enough that we read it as a real difference rather than as sampling noise') and even shows the failure concentration on one side of the distribution (Sec. 4.4). Yet the three comparison rows are measured over 25 trials per task (Table 4 note), with no confidence intervals, and the per-orientation subgroup where the claimed difference is largest has 5 trials per variant. Binomial variance at n=25 gives roughly 2.0 to 2.3 success-rate points (SD) per row; a 12-point gap is about 3.6 SD if independent, so the ablation direction is plausible, but the failure-distribution argument is weaker: in the 90-degree chili group, 1/5 versus 4/5 successes is p approximately 0.2 by Fisher's exact test, and the 'train edge' coverage argument (Fig. 5) is invoked without a controlled test of that subgroup. The gap could also arise from the full grid's longer effective optimization (Sec. 4.4 extends the full grid alone to 200k steps while the reduced sets level off earlier), so the comparison conflates objective count with training schedule. The flow labels from frozen Track4World and the MotionCrafter 4D VAE (Sec. 3.2, Sec. 5) are acknowledged as noisy; if the model learned the tracker's errors rather than true 3D motion, the measured gains would not conceptually support the 4D-geometry claim, but this is secondary because the real-robot ablation attributes gains to the flow objectives themselves.

The reader's weakest_assumption concerns unverified frozen flow pipelines (Track4World pseudo-labels, MotionCrafter 4D VAE compression). That is a real limitation, and the paper itself acknowledges tracker noise (Sec. 5). However, the most load-bearing point for the central claim is not label quality alone: it is that the only evidence isolating the flow objectives' contribution is the real-robot ablation (Table 4), and that evidence is statistically underpowered as reported (25 trials per row, 5 trials in the highlighted subgroup, no confidence intervals), and partially confounded with training schedule because only the full grid is extended to 200k steps (Sec. 4.4). The simulation ablation says the objective sets differ 'mainly by training speed' (Sec. 4.4), so by the paper's own measure simulation does not separate flow value from optimization budget. On LIBERO-Plus, the claimed robustness advantage is reported at best-checkpoint values (87.65%) and the comparison mixes best-evaluated with fixed-schedule baselines, though LIBERO-Plus is a secondary claim. I partially agree with the reader: I agree the frozen-pipeline adequacy is unverified, but the 25-trial real-robot rows plus schedule confounding are the most load-bearing weakness, because the simulation evidence explicitly does not separate objective value from training speed. The verdict remains CONDITIONAL rather than REJECT: the direction of the effect is consistent across tasks, the paper discloses trial counts and the schedule difference, and a targeted re-run could settle the concern. No ad hominem; the critique concerns statistical power and protocol design, not integrity.

The referee's report

If the central claim holds, the paper makes a useful contribution: it is among the first to treat scene-level 3D flow as a jointly denoised, self-supervised modality inside a policy rather than as an external module or inference-time prior, and the ten-objective role grid is a clean way to organize multi-task training across action, vision, and geometry. The simulation evidence on RoboCasa-GR1 is comparatively strong: the repeated-seed evaluation over 11 seeds with standard deviations, the mid-training versus no-mid-training comparison starting from the same base checkpoint, and the task-stratified gain analysis all support the direction of the claim. The paper also makes reasonable use of standard benchmarks (LIBERO, LIBERO-Plus) and reports an objective ladder on the real robot. The main weaknesses are concentrated in the real-robot ablation, where the sample sizes are small and the comparison is not schedule-matched, and in the lack of direct verification that the frozen pseudo-label pipeline produces an accurate 4D geometry signal rather than an auxiliary-task regularizer.

Pith's reading of the paper (a machine's summary, not the authors')

Robot demonstrations record what the robot saw and what it did, but not how the scene moved in 3D in between, and this paper argues that this missing 4D geometry is learnable and worth supervising directly. UniFlowWAM promotes scene-level 3D flow to a third modality on par with vision and action inside one diffusion transformer with a dedicated expert per modality, and a ten-objective role grid assigns each modality the role of target, condition, or absent for every training task. The load-bearing claim is that flow's own supervision improves the action policy: on RoboCasa-GR1 success rises from 52.11% for the released base model to 56.38% with mid-training, and on a real bimanual robot the flow objectives account for a 12-to-18-point gap over policy-only post-training. The gains concentrate where the base model is weakest and where demonstrations are sparse, and the deployed policy still takes no flow input at test time, so whatever flow teaches arrives through the shared representation rather than through an inference-time module.

What carries the argument. The load-bearing mechanism is a three-expert multimodal diffusion transformer in which visual features, 3D flow latents, and action chunks each have a private expert (own parameters, normalization, and modulation) and meet only through joint self-attention over a shared MM-RoPE position frame. Scene-level 3D flow is defined as per-frame 3D point maps plus per-frame displacements, produced offline by a frozen dense 3D point tracker and compressed by a frozen 4D VAE into a latent that the flow expert denoises. Training is organized by a ten-objective role grid: each objective gives every modality exactly one of three roles, namely target (noised and supervised), condition (clean input fixed at the clean end), or absent (learned placeholder visible to attention), so one model covers policy, forward and inverse dynamics, and visual or flow prediction. The loss is the flow-matching velocity loss applied to target slots only, with a task-balanced sampler keeping every objective present in each batch.

Core claim. The paper's central claim is that a world-action model which treats scene-level 3D flow as a co-equal modality, supervised as its own target and usable as a conditioning signal, learns better actions than the same model trained only on vision and actions. The evidence is the ablation ladder on the real robot: policy-only post-training scores 76, 76, and 72 on the three tasks while the full ten-objective grid scores 94, 90, and 84, with the 3D flow objectives adding 10 points on pen manipulation and 8 on towel folding relative to the no-flow variant. The failures of policy-only training concentrate at the edges of the region covered by demonstrations, which the authors read as flow supervision injecting 4D geometric information that the visual and action streams do not carry. Because deployment runs the direct-policy objective, in which the flow stream holds learned placeholders, the reported differences come entirely from training, not from the inference path. The authors also report that mid-training the flow expert on roughly 1,200 hours of demonstrations adds 4.27 points on RoboCasa-GR1 over the released base model, with the largest gains on the tasks where the base model is weakest.

The premise it rests on. The flow signal comes entirely from frozen external models, namely Track4World's pseudo-labels, MotionCrafter's 4D VAE compression, and SAM-based foreground masks, and while the paper acknowledges that these labels contain temporal inconsistencies, it does not independently verify that they teach true 3D motion rather than tracker noise.

If the paper is right (Pith's inferences, not the authors')

Flow supervision, not flow input, is what helps: the policy deploys with learned placeholders in the flow stream, so the reported gains cost nothing at inference and need no tracker or VAE at test time.

The improvement concentrates where the base model is weakest: on the 10 RoboCasa tasks below 50% success, the mid-trained model improves on all 10 by 6.47 points on average, suggesting flow supervision helps most where vision-action data alone is thinnest.

The ten-objective grid keeps in-distribution performance intact while improving out-of-distribution behavior: LIBERO stays at 98.1% and zero-shot LIBERO-Plus reaches 87.7%, with the largest margin under camera perturbations (96.9% versus 76.6% for the strongest baseline).

A single model trained on the grid is at once a policy, a forward and inverse dynamics model, and a flow and visual predictor, so data without action labels or without future frames can in principle contribute to the same weights that serve as the policy.

The documented failures of alternative designs delimit the design space: flow must enter as its own latent modality with its own supervision rather than as re-coded image channels or a gated extra, since the image-VAE variant collapsed to an average flow and the gating variant trailed by 3.3 points.

Editorial inference: the concentration of real-robot failures at the boundary of the demonstrated region suggests flow objectives act as a geometric regularizer rather than as a source of new skills; the testable prediction is that flow supervision matters most under occlusion, deformable objects, and near-edge placements, and least for repetitive primitives like chili transfer.

Editorial inference: because predicted flow latents decode back to point maps and scene flow, the trained model is a latent forward model that could be used for plan verification, checking whether an imagined action sequence produces the scene motion the policy expects, or for monitoring whether the executed action actually moved the scene as predicted.

Editorial inference: the paper's stated limitation that flow pretraining used only about 1,200 hours against the much larger corpus used for the base model implies the reported numbers are a lower bound; joint large-scale pretraining of the flow expert is the direct next experiment, and a simulator comparison with ground-truth flow would separate label quality from the architecture's benefit.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Robot demonstrations record what the robot saw and what it did. They do not specify how the 3D world changed in between. Scene-level 3D flow describes it by tracking 3D scene points, providing the world representation with 4D geometry. Prior work has used flow to improve robot policies. Flow served as an intermediate prediction that a separate model converts into actions, as a prior injected at inference time, or as an extra output learned together with the action. In all of these designs, flow stays outside the policy, as a separate module or as an extra prediction head. We introduce UniFlowWAM, a Unified World-Action Model that makes 3D scene flow a third modality on par with vision and action. Its world representation therefore covers both visual semantics, through future visual features, and 4D geometry, through 3D flow, and flow receives its own supervision. UniFlowWAM is a diffusion transformer with one expert per modality, and the three experts exchange information through joint attention. We train UniFlowWAM under a ten-objective role grid, where each objective assigns every modality exactly one of three roles, namely target, condition, or absent. The same model thus serves as a policy, as forward and inverse dynamics, and as a predictor of future visual features or 3D flow. We evaluate UniFlowWAM on RoboCasa-GR1, LIBERO, and LIBERO-Plus in simulation, and on four real-world bimanual manipulation tasks. On RoboCasa-GR1, UniFlowWAM raises the success rate of its base model without flow from 52.1% to 56.4%. It preserves in-distribution success on LIBERO at 98.1% and reaches 87.7% zero-shot on LIBERO-Plus. On the real robot, it achieves 87.0% average success, compared with 68.5% for the strongest baseline.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-23

The authors' main claim, as the classifier picked it out 0.49

“On the real robot, it achieves 87.0% average success, compared with 68.5% for the strongest baseline.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.11 · states limitation 0.10 · human participants 0.04 · controlled comparison 0.97 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.96 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Robot demonstrations record what the robot saw and what it did. background 0.99 0.10 0.01 asserted 1.00
They do not specify how the 3D world changed in between. gap 0.94 0.11 0.02 asserted 1.00
Scene-level 3D flow describes it by tracking 3D scene points, providing the world representation with 4D geometry. background 0.91 0.11 0.05 asserted 1.00
Prior work has used flow to improve robot policies. background 1.00 0.03 0.02 asserted 1.00
Flow served as an intermediate prediction that a separate model converts into actions, as a prior injected at inference time, or as an extra output learned together with the action. background 1.00 0.03 0.02 asserted 1.00
In all of these designs, flow stays outside the policy, as a separate module or as an extra prediction head. gap 0.58 0.06 0.05 asserted 1.00
We introduce UniFlowWAM, a Unified World-Action Model that makes 3D scene flow a third modality on par with vision and action. contribution 1.00 0.91 0.30 asserted 1.00
Its world representation therefore covers both visual semantics, through future visual features, and 4D geometry, through 3D flow, and flow receives its own supervision. method 0.52 0.87 0.18 asserted 1.00
UniFlowWAM is a diffusion transformer with one expert per modality, and the three experts exchange information through joint attention. method 0.98 0.86 0.06 asserted 1.00
We train UniFlowWAM under a ten-objective role grid, where each objective assigns every modality exactly one of three roles, namely target, condition, or absent. method 0.98 0.95 0.05 asserted 1.00
The same model thus serves as a policy, as forward and inverse dynamics, and as a predictor of future visual features or 3D flow. method 0.55 0.83 0.18 asserted 1.00
We evaluate UniFlowWAM on RoboCasa-GR1, LIBERO, and LIBERO-Plus in simulation, and on four real-world bimanual manipulation tasks. method 0.63 0.96 0.04 asserted 0.98
On RoboCasa-GR1, UniFlowWAM raises the success rate of its base model without flow from 52.1% to 56.4%. result 1.00 0.97 0.11 asserted 1.00
It preserves in-distribution success on LIBERO at 98.1% and reaches 87.7% zero-shot on LIBERO-Plus. result 1.00 0.96 0.08 asserted 1.00
On the real robot, it achieves 87.0% average success, compared with 68.5% for the strongest baseline. result 1.00 0.97 0.19 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.90

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-23: 1 request, 10,796 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (5)
Figure 1 of the paper
Figure 1. UniFlowWAM overview: UniFlowWAM couples visual features, scene-level 3D flow, and actions in one three-expert diffusion transformer, denoising the three jointly so that 3D flow carries scene displacement into the space where actions are executed. A grid of ten objectives assigns each modality the role of target, condition, or absent, so a single model covers policy, dynamics, and prediction, yet deploys as a direct policy that needs no flow at test time. makes the effects of robot commands on the scene explicit: it describes how each scene point is displaced in 3D over time,…
Figure 2 of the paper
Figure 2. Network architecture. Actions, scene-level 3D flow, and future visual features are three modalities of one diffusion transformer, each processed by its own expert; the experts exchange information only through joint self-attention. The objective decides which modalities are noised targets and which are clean conditions. Predicted flow latents are decoded by the 4D VAE into point maps and scene flow. and plans actions from it. Im2Flow2Act [38] and PointAction [31] predict flow or point trajectories and convert them into actions with a separate module, so the predictor is trained…
Figure 3 of the paper
Figure 3. Real-world manipulation tasks. Instructions and scored milestones for four tasks: placing a cabbage in a pot and closing the lid, transferring two chilies onto a plate, placing two pens in a drawer and closing it, and folding a towel. Demonstrations. We consider four tasks that differ in what they demand of the two arms (
Figure 4 of the paper
Figure 4. Results of real-world manipulation on ALOHA. Solid bars give the full-task success rate; the lighter dashed segment extends to the process score, the fraction of ordered milestones completed (
Figure 5 of the paper
Figure 5. Demonstration coverage on two real-robot tasks. Left: chili transfer, with coverage measured over the chili’s rotation angle: 0◦ is lengthwise, ±90◦ sideways. Right: pen manipulation, with coverage measured over the pen’s placement in the top camera. The red box spans the initial pen positions of all 49 demonstrations; the blue band is its boundary, where demonstrations are naturally sparse. earns credit for the later one. Both metrics are averaged over trials, then over tasks. Task performance. UniFlowWAM reaches 87.0% average success with a process score of 92.3, against 68.5% and…

Figures are the authors', reproduced from the paper.

References (51)
  1. [1] H.Bi, H.Tan, S.Xie, Z.Wang, S.Huang, H.Liu, R.Zhao, Y.Feng, C.Xiang, Y.Rong, etal. Motus: Aunified latent action world model. InCVPR, 2026 on Pith
  2. [2] J.Bjorck,F.Castañeda,N.Cherniadev,X.Da,R.Ding,L.Fan,Y.Fang,D.Fox,F.Hu,S.Huang,etal. GR00T N1: An open foundation model for generalist humanoid robots.arXiv:2503.14734, 2025 on PithPith's reviewarXiv
  3. [3] K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al. 𝜋0: A vision-language-action flow model for general robot control. InRSS, 2025. 12 UniFlowWAM on Pith
  4. [4] Q.Bu,Y.Yang,J.Cai,S.Gao,G.Ren,M.Yao,P.Luo,andH.Li. LearningtoActAnywherewithTask-centric Latent Actions. InRSS, 2025 on Pith
  5. [5] Y. Cao, S. Liu, C. Feng, M. Zhang, X. Lu, A. Owens, and K. Fang. Unified motion-action modeling for heterogeneous robot learning.arXiv:2606.16917, 2026 on PithPith's reviewarXiv
  6. [6] S. Community. StarVLA: A Lego-like codebase for vision-language-action model developing. arXiv:2604.05014, 2026 on PithPith's reviewarXiv
  7. [7] P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InICML, 2024 on Pith
  8. [8] H. Fang, J. Duan, D. Clay, S. Wang, S. Liu, W. Huang, X. Fan, W.-C. Tsai, S. Chen, Y. R. Wang, et al. MolmoAct2: Action reasoning models for real-world deployment. InCoRL, 2026 on Pith
  9. [9] S.Fei,S.Wang,J.Shi,Z.Dai,J.Cai,P.Qian,L.Ji,X.He,S.Zhang,Z.Fei,etal. LIBERO-Plus: AProgressive Robustness Benchmark for Visual-Language-Action Models. InCVPR, 2026 on Pith
  10. [10] J. Guo, Q. Li, P. Li, Z. Chen, N. Sun, Y. Su, H. Wang, Y. Zhang, X. Li, and H. Liu. Unified 4D world action modeling from video priors with asynchronous denoising.arXiv:2604.26694, 2026 on PithPith's reviewarXiv
  11. [11] C. Hou, K. Wu, J. Liu, Z. Che, D. Wu, F. Liao, G. Li, J. He, Q. Feng, Z. Jin, et al. RoboMIND 2.0: A multimodal,bimanualmobilemanipulationdatasetforgeneralizableembodiedintelligence.arXiv:2512.24653, 2025 on PitharXiv
  12. [12] Y. Hu, Y. Guo, P. Wang, X. Chen, Y.-J. Wang, J. Zhang, K. Sreenath, C. Lu, and J. Chen. Video prediction policy: A generalist robot policy with predictive visual representations. InICML, 2025 on Pith
  13. [13] 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. InCoRL, 2024 on Pith
  14. [14] M. J. Kim, C. Finn, and P. Liang. Fine-tuning vision-language-action models: Optimizing speed and success. InRSS, 2025 on Pith
  15. [15] J. Lee, J. Duan, H. Fang, Y. Deng, S. Liu, B. Li, B. Fang, J. Zhang, Y. R. Wang, S. Lee, et al. MolmoAct: Action reasoning models that can reason in space. InICRA, 2026 on Pith
  16. [16] H.Li,L.Sun,Y.Hu,D.Ta,J.Barry,G.Konidaris,andJ.Fu. NovaFlow: Zero-shotmanipulationviaactionable flow from generated videos. InICRA, 2026 on Pith
  17. [17] S. Li, Y. Gao, D. Sadigh, and S. Song. Unified Video Action Model. InRSS, 2025 on Pith
  18. [18] B. Lin, Y. Xu, X. Zhan, H. Fang, J. Tian, F.-C. Zhang, Y.-L. Li, C. Lu, and L. Yang. ChronoFlow-Policy: Unifying past-current-future interaction flow in visuomotor policy learning. InECCV, 2026 on Pith
  19. [19] Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. In ICLR, 2023 on Pith
  20. [20] B. Liu, Y. Zhu, C. Gao, Y. Feng, Q. Liu, Y. Zhu, and P. Stone. LIBERO: Benchmarking knowledge transfer for lifelong robot learning. InNeurIPS, 2023 on Pith
show 31 more
  1. [21] J. Lu, W. Xiong, J. Deng, P. Li, T. Huang, Z. Dou, C. Lin, S.-K. Yeung, and Y. Liu. TrackingWorld: World-centric monocular 3D tracking of almost all pixels. InNeurIPS, 2025 on Pith
  2. [22] J. Lu, J. Xu, W. Hu, R. Zhu, C. Zhao, S.-K. Yeung, Y. Shan, and Y. Liu. Track4World: Feedforward world-centric dense 3D tracking of all pixels. InECCV, 2026 on Pith
  3. [23] H.Luo,W.Zhang,Y.Feng,S.Zheng,H.Xu,C.Xu,Z.Xi,Y.Fu,andZ.Lu. Being-H0.7: Alatentworld-action model from egocentric videos.arXiv:2605.00078, 2026 on PithPith's reviewarXiv
  4. [24] J. Lyu, K. Liu, X. Zhang, H. Liao, Y. Feng, W. Zhu, T. Shen, J. Chen, J. Zhang, Y. Dong, et al. LDA-1B: Scaling latent dynamics action model via universal embodied data ingestion. InRSS, 2026 on Pith
  5. [25] S. Nasiriany, A. Maddukuri, L. Zhang, A. Parikh, A. Lo, A. Joshi, A. Mandlekar, and Y. Zhu. RoboCasa: Large-Scale Simulation of Household Tasks for Generalist Robots. InRSS, 2024 on Pith
  6. [26] S. Noh, D. Nam, K. Kim, G. Lee, Y. Yu, R. Kang, and K. Lee. 3D flow diffusion policy: Visuomotor policy learning via generating flow in 3D space.arXiv:2509.18676, 2025 on PitharXiv
  7. [27] Physical Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, et al.𝜋0.5: a vision-language- action model with open-world generalization. InCoRL, 2025. 13 UniFlowWAM on Pith
  8. [28] D. Qu, H. Song, Q. Chen, Y. Yao, X. Ye, Y. Ding, Z. Wang, J. Gu, B. Zhao, D. Wang, et al. SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Models. InRSS, 2025 on Pith
  9. [29] O. Siméoni, H. V. Vo, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V. Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, et al. DINOv3.arXiv:2508.10104, 2025 on PithPith's reviewarXiv
  10. [30] Z. Teed and J. Deng. RAFT: Recurrent all-pairs field transforms for optical flow. InECCV, 2020 on Pith
  11. [31] M. Tong, H. Jiang, Q. Feng, L. Liu, and J. Gu. PointAction: 3D points as universal action representations for robot control.arXiv:2606.03943, 2026 on PithPith's reviewarXiv
  12. [32] C. Wang, X. Wang, B. Lin, J. Tian, F. Zhang, C. Lu, and L. Yang. Track4Action: Distilling world-centric 3D tracker into vision-language-action policies.arXiv:2608.03727, 2026 on PithPith's reviewarXiv
  13. [33] J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny. VGGT: Visual geometry grounded transformer. InCVPR, 2025 on Pith
  14. [34] X. Wang, C. Wang, Y. Xu, M. Ye, F. Zhang, J. Tian, X. Zhan, L. Zhu, C. Lu, and L. Yang. LaMP: Learning Vision-Language-Action Policy with 3D Scene Flow as Latent Motion Prior. InECCV, 2026 on Pith
  15. [35] Y. Wang, S. Huang, M. Li, C. Zhang, J. Liang, W. Jin, Y. Chen, X. Chi, D. Zhou, Q. Yu, et al. OpenWAM: An open, modular exploration towards systematic world-action model pretraining.arXiv:2609.07398, 2026 on PithPith's reviewarXiv
  16. [36] S. Wu, X. Liu, S. Xie, P. Wang, X. Li, B. Yang, Z. Li, K. Zhu, H. Wu, Y. Liu, et al. RoboCOIN: An open-sourced bimanual robotic data collection for integrated manipulation.arXiv:2511.17441, 2025 on PithPith's reviewarXiv
  17. [37] Y. Xiao, J. Wang, N. Xue, N. Karaev, Y. Makarov, B. Kang, X. Zhu, H. Bao, Y. Shen, and X. Zhou. SpatialTrackerV2: Advancing 3D Point Tracking with Explicit Camera Motion. InICCV, 2025 on Pith
  18. [38] M. Xu, Z. Xu, Y. Xu, C. Chi, G. Wetzstein, M. Veloso, and S. Song. Flow as the cross-domain manipulation interface. InCoRL, 2024 on Pith
  19. [39] G. Yan, J. Liu, Y. Fan, L. Cai, M. Liao, J. Zhang, and D. Fox. Flex-𝜋: A multi-stream world-action model with compute flexibility.arXiv:2608.10860, 2026 on PithPith's reviewarXiv
  20. [40] L. Yang, W. Song, X. Wang, P. Sheng, Z. Fang, Z. Zhou, J. He, H. Yan, J. Chen, N. Sun, et al. 4D-WAM: Infusing spatiotemporal awareness into world action models through trajectory fields.arXiv:2608.08023, 2026 on PithPith's reviewarXiv
  21. [41] J.Ye,N.Gao,S.Yang,J.Zheng,Z.Wang,Y.Chen,P.Chen,Y.Chen,S.Liu,andJ.Jia. StarVLA- 𝛼: Reducing complexity in vision-language-action systems. InECCV, 2026 on Pith
  22. [42] C. Yuan, C. Wen, T. Zhang, and Y. Gao. General flow as foundation affordance for scalable robot learning. In CoRL, 2024 on Pith
  23. [43] H. Yuan, W. Chen, J. Cen, H. Yu, J. Liang, S. Chang, Z. Lin, T. Feng, P. Liu, J. Xing, et al. Lumos-1: On Autoregressive Video Generation with Discrete Diffusion from a Unified Model Perspective. InICLR, 2026 on Pith
  24. [44] T. Yuan, Z.Dong, Y. Liu, and H.Zhao. Fast-WAM: Doworld action modelsneed test-timefuture imagination? arXiv:2603.16666, 2026 on PithPith's reviewarXiv
  25. [45] W. Zhang, H. Liu, Z. Qi, Y. Wang, X. Yu, J. Zhang, R. Dong, J. He, H. Wang, Z. Zhang, et al. DreamVLA: A vision-language-action model dreamed with comprehensive world knowledge. InNeurIPS, 2025 on Pith
  26. [46] R. Zheng, Y. Liang, S. Huang, J. Gao, H. Daumé III, A. Kolobov, F. Huang, and J. Yang. TraceVLA: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies. InICLR, 2025 on Pith
  27. [47] R. Zheng, J. Wang, S. Reed, J. Bjorck, Y. Fang, F. Hu, J. Jang, K. Kundalia, Z. Lin, L. Magne, et al. FLARE: Robot learning with implicit world modeling. InCoRL, 2025 on Pith
  28. [48] Z. Zhong, H. Yan, J. Li, X. Liu, X. Gong, T. Zhang, W. Song, J. Chen, X. Zheng, H. Wang, et al. FlowVLA: Visual chain of thought-based motion reasoning for vision-language-action models.arXiv:2508.18269, 2025 on PitharXiv
  29. [49] G. Zhou, H. Pan, Y. LeCun, and L. Pinto. DINO-WM: World models on pre-trained visual features enable zero-shot planning. InICML, 2025 on Pith
  30. [50] C. Zhu, R. Yu, S. Feng, B. Burchfiel, P. Shah, and A. Gupta. Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets. InRSS, 2025 on Pith
  31. [51] R. Zhu, J. Lu, W. Hu, X. Han, J. Cai, Y. Shan, and C. Zheng. MotionCrafter: Dense geometry and motion reconstruction with a 4D VAE. InCVPR, 2026. 14 on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment