Pith.

Robotics · arXiv:2610.18697 v2 · submitted 13 Oct 2026

UniW(A)M: A Unified Mixture-of-Experts Framework for World and World-Action Models

Jingbo Xu, Jianjie Fang, Zhuohang Li, Zihao Song, Zhaolu Wang, Xvyuan Liu, Xin Zhang, Haisheng Su

show 4 more authors

Yinzhou Tang, Chen Gao, Wei Wu, Yong Li

The paper says

“World models (WMs) and world-action models (WAMs) focus on future world prediction and executable action generation, respectively.”

more of the abstract

Although these models learn shared physical dynamics and interaction semantics from embodied trajectories, they are typically trained in isolation, which restricts knowledge reuse between prediction and control. The central challenge is therefore to develop a unified learning framework that shares these dynamics while preserving the distinct modeling capabilities required by each paradigm. We propose UniW(A)M, a mixture-of-experts world-(action) model built upon diffusion Transformers that unifies action-conditioned world modeling and diverse WAM variants. Shared experts capture common dynamics, while atomic capability experts are reused across interfaces so that each interface keeps its own conditional pathway. At deployment, a plug-in adaptive interface router (PAIR) learns from a small subset of target-task chunk scenarios to select a suitable WAM interface for each execution chunk, balancing task success and execution efficiency while keeping the backbone frozen. We evaluate UniW(A)M on LIBERO, RoboTwin 2.0, and three real-world manipulation tasks, where it achieves 88.3% average real-world success versus 76.7% for the strongest baseline. On WorldArena, it leads action-conditioned world modeling on all three tasks. AC-WM co-training contributes to action generation on every action interface and PAIR raises OOD success from 63.3% to 76.7%.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 14 Oct 2026 · of version 1 of the paper; version 2 is newer · 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. The real contribution is the joint-training design and its clean ablations; the headline 88.3% real-world figure does not survive a difference test. This line is wrong

3 major objections · 4 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 references say.

The paper says, at Abstract; Section 4.2; Table 2; Table 8; Appendix C.1

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.60):

“One additional success changes a task’s observed rate by five percentage points, so small differences in point estimates should not be interpreted as stable performance rankings.”

Pith says: Objection 1 of 3 · major

The headline comparison '88.3% average real-world success versus 76.7% for the strongest baseline' is asserted in the abstract and Section 4.2 without a difference test, and the counts reported in Table 8 do not support it at conventional significance levels.

more of objection 1

Per-task Fisher exact tests on IDM-WAM versus WorldScape Policy 2.0 give p about 0.41 (cloth folding, 18/20 vs 15/20), p about 0.70 (box folding, 17/20 vs 15/20), and p about 0.66 (tabletop cleaning, 18/20 vs 16/20); even the pooling that Appendix C.1 explicitly declines, 53/60 versus 46/60, gives two-sided p about 0.09. The issue is compounded because 88.3% is the best of four correlated interface estimates (no multiple-comparison control) and the margins of the weaker interfaces (for example, Action-Only-WAM 80.0% vs 76.7%) are plainly within noise. The disclosure of marginal Wilson intervals in Appendix C.1 is commendable, and the authors' warning that one additional success changes a rate by five percentage points is appropriately cautious; what remains unresolved is that the abstract and Section 4.2 present the 11.7-point gain as an established result. Please add a test that respects the three-task block structure (for example, a stratified permutation test or a mixed-effects logistic regression), or qualify the headline with its non-significance.

The paper says, at Abstract; Section 4.3; Table 4; Table 8; Eq. (1); Eq. (13); Appendix B.1

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The claim that PAIR raises OOD success from 63.3% to 76.7% rests on the same 20-trial-per-cell structure and is likewise unsupported by a difference test. From Table 8, the best standalone interface achieves 38/60 and PAIR achieves 46/60; the most favorable pooling gives two-sided p about 0.11, and per-task Fisher exact tests are non-significant (the smallest is about p=0.07 for cloth folding, 15/20 vs 13/20).

more of objection 2

The paper states the relevant caveats (marginal intervals, no training-run variance, no pooling) in Appendix C.1, so the unresolved step is the unqualified causal phrasing. Two further points matter. First, the routing target in Eq. (1) depends on the temperature tau and on the near-best threshold delta=0.01 in Eq. (13), with no sensitivity analysis reported for either; because the observed gain is small relative to sampling noise, a delta/tau sweep is needed to show the result is not confined to a narrow regime. Second, PAIR is the only method that consumes the OOD demonstrations (2,000 chunk scenarios per task, Appendix B.1), so the comparison against zero-shot standalone interfaces does not isolate chunk-level switching from the mere availability of target-domain data; a control that selects a single interface per task with the same MAE labels would help separate these effects.

The paper says, at Abstract; Section 4.2; Table 3; Tables 12-13

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The abstract's claim that UniW(A)M leads action-conditioned world modeling on all three tasks rests on Task 2 (Folding Box), where UniW(A)M's aggregate EWMScore-P is 55.89 versus 55.48 for CtrlWorld (Table 3), a 0.41-point margin, and the component scores in Tables 12-13 are mixed (UniW(A)M trails Cosmos-Predict 2.5 on InstructionFollowing, 84.0 vs 95.0).

more of objection 3

No uncertainty is reported for the aggregate metric, so 'leads on all three tasks' overstates what the data establish. Please report how EWMScore-P is aggregated and any variance across the evaluated clips, or soften the wording.

4 minor objections

The paper says, at Tables 2 and 11; References

Pith says: minor objection 1 of 4

WorldScape Policy 2.0 (Su et al., 2026), the strongest baseline in Tables 2 and 11, shares six co-authors with this manuscript (Jianjie Fang, Xin Zhang, Haisheng Su, Chen Gao, Wei Wu, Yong Li); please disclose this overlap when presenting the point-estimate gap, so that readers do not mistake the comparison for an external one.

The paper says, at Reproducibility statement; Appendices A.3-A.4

Pith says: minor objection 2 of 4

The claimed independence of the four interfaces is an implementation-level property that depends on teacher-forcing copy separation, per-interface KV caches, and cache clearing at interface changes. The paper describes these mechanisms carefully, but with no code release the property cannot be checked; a targeted leakage test (for example, verifying that each interface's outputs are invariant to perturbations of tokens the mask should hide) would materially increase confidence in the conditional-pathway claim.

show 2 more

The paper says, at Equation (11)

Pith says: minor objection 3 of 4

The normalization in the MAE definition L(i)_MAE is ambiguous as typeset ('1kda'); please verify the intended denominator (presumably 1/(k*d_a)).

The paper says, at Section 4.2; Table 10

Pith says: minor objection 4 of 4

The LIBERO paragraph highlights the 99.2% Spatial score, but UniW(A)M's 97.5% average is below two cited WAM baselines (Motus 97.7 and LingBot-VA 98.5 in Table 10); one sentence of context would prevent the paragraph from reading as a state-of-the-art claim.

What would settle it

Run a two-sided, task-stratified permutation test on the Table 8 counts, respecting the per-task 20-trial structure: IDM-WAM (18,17,18) vs WorldScape Policy 2.0 (15,15,16) for the standard setting, and PAIR (15,14,17) vs the best standalone interface per task (13,11,15) for OOD; alternatively, fit a fixed-effects logistic regression of success on method with task strata. If either p ≥ 0.05, as the pooled approximations suggest, the abstract's comparative claims must be reported as within-noise, or supported by a pre-registered evaluation with n ≥ 50 trials per cell and a stated difference test; if both p < 0.05, the numerical headline survives and this concern is refuted.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 14 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 2 min 6 s · 29,639 tokens read, 18,730 written
  3. Editor wrote the verdict above 49 s · 30,768 tokens read, 6,199 written
  4. Referee major revision: 3 major objections, 4 minor 3 min 0 s · 30,183 tokens read, 28,395 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

Two things you should know before reading this. The design work is genuinely good: five video-action interfaces trained jointly in one MoE diffusion transformer, with CATR routing token groups to capability-matched experts. The key ablation is parameter-matched and clean — sample-level routing collapses the action-only interfaces to 10% average success while CATR holds all four interfaces between 80% and 88.3% — and the AC-WM co-training ablation shows large drops on every interface when the world-modeling paradigm is removed, even on the two interfaces that never predict video at deployment. The RoboTwin C2R margin over pi0.5 (+20.6 points, 100 trials per task, official leaderboard protocol) is the most solid external evidence in the paper.

The numbers that headline the abstract do not survive a difference test. The stress-test arithmetic checks out, and it is the aggregation most favorable to the authors: IDM-WAM versus WorldScape Policy 2.0 is 53/60 versus 46/60 (two-sided p about 0.09), and PAIR's OOD gain is 46/60 versus 38/60 (p about 0.11). Every per-task Wilson interval overlaps its baseline. Credit where it is due: Appendix C.1 discloses the intervals and explicitly declines difference tests — but then the abstract and Section 4.2 assert the 11.7-point gain flatly. The strongest baseline also shares six authors with this paper, and the overlap is not disclosed. Those two things need fixing: either difference tests or run-level variance, and a provenance note on the baseline.

The concerns I would not press as hard: the interface-mask leakage worry is not demonstrated. The appendix specifies teacher forcing and cache handling carefully enough that the conditional-dependency claim is checkable rather than obviously broken. And PAIR's supervision — labels derived from the frozen model's own MAE against target-task demonstrations — is a reasonable design choice, not a circular evaluation, because OOD success is measured in fresh physical trials. The no-code, no-data situation is a real problem for a paper whose flagship evidence is twenty trials per cell.

Net: this is a serious empirical study whose mechanism claims are better supported than its headline. Send it to referees; ask for uncertainty quantification, baseline disclosure, and artifacts. I would not cite the 88.3% number as a result, but I would cite the CATR finding and the five-interface taxonomy.

The strongest objection

The paper's flagship empirical claim is that IDM-WAM reaches 88.3% average real-world success versus 76.7% for the strongest baseline WorldScape Policy 2.0 (abstract; Table 2), and that PAIR raises OOD success from 63.3% to 76.7% (Table 4). Both claims rest on 20 trials per method–task cell, and the reported counts do not distinguish the model from the baseline. Pooling the three cells, the most favorable simple aggregation: IDM-WAM is 53/60 vs 46/60 (normal-approximation z≈1.70, two-sided p≈0.09); PAIR is 46/60 vs 38/60 (z≈1.61, p≈0.11). Every per-task Wilson interval in Table 8 overlaps its baseline (e.g., cloth folding 18/20, [69.9,97.2] versus 15/20, [53.1,88.8]). These calculations are already favorable to the authors: the headline 88.3% is the best of four correlated interface estimates, no training-run variance is reported, and no multiple-comparison control is applied. Appendix C.1 discloses marginal intervals and explicitly declines difference tests, but the abstract and Section 4.2 assert the 11.7-point gain without qualification; the unresolved step is whether the flagship comparative claims are real, and the test is cheap because the counts are in the paper. It is also unstated that WorldScape Policy 2.0 shares six authors with this paper (Su et al. 2026 versus the UniW(A)M author list), so the point-estimate gap is an in-group comparison whose real-world numbers may have been produced by the same team. In fairness, the mechanism claims are better supported: the CATR ablation is parameter-matched with large gaps, the AC-WM co-training ablation drops every interface, and the RoboTwin C2R margin (+20.6 points, 100 trials per task) is solid. The concern targets specifically the real-world superiority numbers that headline the abstract.

I read the central claim as empirical: joint training with CATR improves every WAM interface and beats the strongest baseline by 11.7 points on real-world tasks, and PAIR gains 13.3 points OOD with the backbone frozen. The least secure premise is that these headline magnitudes are real rather than noise. With 20 trials per cell, no difference tests, best-of-four interface selection for the headline, single training runs, and an in-group strongest baseline, the reported data are consistent with no true gap; the paper's own caution in C.1 that small point-estimate differences should not be read as stable rankings is not reflected in the abstract's unqualified 88.3% versus 76.7%. The authors disclose the intervals but decline the test that would resolve the claim, so the concern is about what remains asserted, not about what is disclosed.

I considered the reader's designated weakest assumption, interface leakage via masks, teacher forcing, and caches. The paper specifies the masks (A.2), teacher-forcing layout (A.3, Table 5), expert eligibility (A.4, Table 6), and cache clearing on interface switches (A.4) in detail, and its own results suggest distinct pathways: OOD success spreads 48.3–63.3 across interfaces, and sample-level routing collapses Action-Centric/Action-Only to 10% while CATR restores them. I found no described mechanism by which leakage necessarily occurs; the residual risk is an implementation bug that code release would settle. That is a verification gap, secondary to the statistical concern, which attacks the claim's numbers directly and is checkable from the paper's own tables.

The framework claims are more robust than the headline: the CATR versus sample-level-routing ablation is parameter-matched with 10% versus 80% gaps; the AC-WM co-training ablation drops every interface; RoboTwin C2R uses 100 trials per task and exceeds pi0.5 by 20.6 points. This is why I keep the reader's CONDITIONAL verdict rather than moving to REJECT: the model is plausibly effective, but the flagship real-world superiority numbers need a stratified difference test, more trials or seed variance, code release, and disclosure of whether the in-group baseline's real-world numbers were re-run by the authors. UNCHANGED means the conditional accept stays, with the conditions sharpened around statistical support for the headline comparisons.

The referee's report

A significant contribution if the results hold, and the mechanism-level evidence is strong. The CATR versus sample-level routing ablation (Table 15, Appendix G.4) is parameter-matched in expert and parameter counts, and its effect sizes are large: pooled counts (Table 9) give z about 7.9 for Action-Centric-WAM and Action-Only-WAM rising from 10.0% to about 80-82% average success. The AC-WM co-training ablation improves every interface-task pair, and the authors honestly note that the four-WAM-only control receives more updates per interface, a confound that opposes their conclusion. The uncertainty reporting in Appendix C.1 (counts plus Wilson intervals) is transparent; the mask formalization in Appendices A.1-A.4 is careful; and the PAIR evaluation uses new physical trials rather than window-level splits, so it is genuinely out-of-distribution. The RoboTwin C2R result (66.60% vs 46.00%, 100 trials per task) provides statistically solid generalization evidence.

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

World models that predict future video and world-action models that generate executable actions are usually trained separately, even though both learn the same physical dynamics from the same embodied trajectories. This paper argues that the two can be unified: one mixture-of-experts diffusion Transformer is jointly trained for action-conditioned world prediction and four different action-generation interfaces, each defined by a visibility mask over future video and action tokens. The payoff, if the claim holds, is that world modeling improves action generation on every interface, including the two that skip future-video prediction at deployment, while prediction quality on the WorldArena benchmark stays ahead of the baselines. On three real-world manipulation tasks the best interface reaches 88.3% average success versus 76.7% for the strongest baseline, and a frozen-backbone router lifts out-of-distribution success from 63.3% to 76.7%.

What carries the argument. The load-bearing machinery is the pairing of video–action visibility masks with capability-aligned expert routing. The masks specify, for each of the five interfaces, which future tokens may attend to which during denoising: mutual visibility for Joint-WAM, one-way video-to-action for IDM-WAM, action-to-video for Action-Centric-WAM, none between future modalities for Action-Only-WAM, and clean supplied actions conditioning video in AC-WM. On top of this, CATR assigns each video or action target group a prescribed eligible expert set (shared-plus-VGM or shared-plus-WM for video, and shared-plus-IDM or shared-plus-VLA for actions) and learns group-level mixture weights only within that set, so specialist computation is aligned with each token's predictive role. Because each expert is copied from the pretrained dense feed-forward network, the MoE replacement preserves the feed-forward output at initialization. PAIR is the deployment-time counterpart: an eight-head classifier on frozen backbone features that selects one of the four WAM interfaces per chunk, trained against MAE-based soft targets with a preference for the faster interface among near-best candidates.

Core claim. The central claim is that shared dynamics and interaction semantics can be reused across prediction and control without sacrificing either capability. UniW(A)M defines five interfaces over the same trajectories — AC-WM and four WAM variants — as conditional pathways fixed by video–action visibility masks, organized around four atomic capabilities (action-free video prediction, direct action generation, action-conditioned video prediction, and video-conditioned action generation). Shared experts capture common dynamics while capability experts specialize and are reused across interfaces; Capability-Aligned Token Routing (CATR) restricts each token group to its eligible experts with learned mixture weights. Joint training makes AC-WM co-training contribute to action generation on every action interface, and the ablations report that sample-level routing collapses the two action-only interfaces to 10% average success while CATR raises them to 81.7% and 80.0%. The paper further claims that a plug-in adaptive interface router (PAIR), trained on 2,000 target-task chunk scenarios with the backbone frozen, selects a WAM interface per execution chunk and raises average out-of-distribution success from 63.3% to 76.7%, landing the system on the Pareto frontier of success versus completion time.

The premise it rests on. The claim stands or falls on whether the interface-specific attention masks, CATR eligibility sets, and cached inference paths enforce exactly the conditional dependencies they describe, so that no interface ever sees future information it is supposed to be blind to during training, teacher forcing, or interface switching; if information leaks, the four WAM interfaces are not distinct conditional pathways and the interpretations of CATR and PAIR lose their basis.

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

AC-WM co-training improves action generation on every WAM interface, including the two that do not predict future video at deployment, so prediction and control objectives can reinforce rather than compete within one backbone.

With identical expert and parameter counts, CATR sustains all four interfaces while sample-level routing collapses the action-only interfaces to 10% average success, indicating that token-group alignment, not expert count, is what preserves interface performance.

Chunk-level interface switching with a frozen backbone raises out-of-distribution success from 63.3% to 76.7%, so deployment-time adaptation need not retrain or fine-tune the shared model.

PAIR's phase-dependent selections, favoring fast action pathways during approach and video-conditioned interfaces during contact and fine manipulation, put the system on the upper-left Pareto frontier of success versus completion time.

The unified model also leads action-conditioned world prediction on all three WorldArena tasks, so the claimed sharing does not trade away prediction quality for action gain.

If the interfaces are genuinely leak-free conditional pathways, the capability decomposition suggests a recipe for extension: a new interface is just a new visibility mask over the same experts, so variants such as language-conditioned or object-centric futures could be added without retraining the backbone.

PAIR's training signal is competence-based, labelling each chunk by which frozen interface predicts the recorded actions best, so the same MAE-targeted router idea transfers to other settings where several frozen decoders or checkpoints compete for the same rollout.

A testable corollary is that PAIR's out-of-distribution gain concentrates in contact-rich chunks; a router variant that switches interfaces only when contact is detected would isolate whether the benefit comes from foresight preference or from chunk-level mixing in general.

A natural stress test of the sharing claim is to shrink the backbone's trajectory data and measure whether the AC-WM co-training benefit grows, which would quantify how much world modeling contributes to the data efficiency of action generation.

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

World models (WMs) and world-action models (WAMs) focus on future world prediction and executable action generation, respectively. Although these models learn shared physical dynamics and interaction semantics from embodied trajectories, they are typically trained in isolation, which restricts knowledge reuse between prediction and control. The central challenge is therefore to develop a unified learning framework that shares these dynamics while preserving the distinct modeling capabilities required by each paradigm. We propose UniW(A)M, a mixture-of-experts world-(action) model built upon diffusion Transformers that unifies action-conditioned world modeling and diverse WAM variants. Shared experts capture common dynamics, while atomic capability experts are reused across interfaces so that each interface keeps its own conditional pathway. At deployment, a plug-in adaptive interface router (PAIR) learns from a small subset of target-task chunk scenarios to select a suitable WAM interface for each execution chunk, balancing task success and execution efficiency while keeping the backbone frozen. We evaluate UniW(A)M on LIBERO, RoboTwin 2.0, and three real-world manipulation tasks, where it achieves 88.3% average real-world success versus 76.7% for the strongest baseline. On WorldArena, it leads action-conditioned world modeling on all three tasks. AC-WM co-training contributes to action generation on every action interface and PAIR raises OOD success from 63.3% to 76.7%.

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 · 2026-10-14

The authors' main claim, as the classifier picked it out 0.93

“We evaluate UniW(A)M on LIBERO, RoboTwin 2.0, and three real-world manipulation tasks, where it achieves 88.3% average real-world success versus 76.7% for the strongest baseline.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.14 · human participants 0.05 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03

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
World models (WMs) and world-action models (WAMs) focus on future world prediction and executable action generation, respectively. background 1.00 0.05 0.02 asserted 1.00
Although these models learn shared physical dynamics and interaction semantics from embodied trajectories, they are typically trained in isolation, which restricts knowledge reuse between prediction and control. gap 0.64 0.07 0.03 asserted 0.98
The central challenge is therefore to develop a unified learning framework that shares these dynamics while preserving the distinct modeling capabilities required by each paradigm. gap 0.89 0.05 0.05 asserted 0.99
We propose UniW(A)M, a mixture-of-experts world-(action) model built upon diffusion Transformers that unifies action-conditioned world modeling and diverse WAM variants. contribution 0.89 0.91 0.12 asserted 1.00
Shared experts capture common dynamics, while atomic capability experts are reused across interfaces so that each interface keeps its own conditional pathway. method 0.98 0.59 0.07 asserted 1.00
At deployment, a plug-in adaptive interface router (PAIR) learns from a small subset of target-task chunk scenarios to select a suitable WAM interface for each execution chunk, balancing task success and execution efficiency while keeping the backbone frozen. method 0.95 0.86 0.08 asserted 0.99
We evaluate UniW(A)M on LIBERO, RoboTwin 2.0, and three real-world manipulation tasks, where it achieves 88.3% average real-world success versus 76.7% for the strongest baseline. result 1.00 0.97 0.40 asserted 1.00
On WorldArena, it leads action-conditioned world modeling on all three tasks. result 1.00 0.93 0.14 asserted 1.00
AC-WM co-training contributes to action generation on every action interface and PAIR raises OOD success from 63.3% to 76.7%. result 0.96 0.92 0.17 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.96

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, 2026-10-14: 1 request, 6,990 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 (19)
Figure 1 of the paper
Figure 1. Overview of UniW(A)M. The model unifies AC-WM and four WAM paradigms over
Figure 2 of the paper
Figure 2. Architecture of UniW(A)M. (a) Stage I: Unified learning. Five interfaces share a diffusion
Figure 3 of the paper
Figure 3. Real-world tasks: cloth folding, box folding, and tabletop cleaning.
Figure 4 of the paper
Figure 4. RoboTwin 2.0 success rates with clean-only training.
Figure 5 of the paper
Figure 5. Inference efficiency and task-level Pareto analysis of UniW(A)M.
Figure 6 of the paper
Figure 6. Joint-WAM visibility masks. Future video and action tokens attend to both modalities,
Figure 7 of the paper
Figure 7. IDM-WAM visibility masks. Action queries can access future video, while video queries
Figure 8 of the paper
Figure 8. Action-Centric-WAM visibility masks. During training, future video queries access ac
Figure 9 of the paper
Figure 9. Action-Only-WAM visibility masks. Neither future modality attends to the other during
Figure 10 of the paper
Figure 10. AC-WM visibility masks. Future video queries attend to the supplied action sequence
Figure 11 of the paper
Figure 11. Expert activation across five video-action interfaces. Shared experts are active for every
Figure 12 of the paper
Figure 12. Phase-conditioned PAIR interface selection across three real-world tasks. Each panel
Figure 13 of the paper
Figure 13. Real-world visualizations of Folding Clothes for the four WAM interfaces. The se
Figure 14 of the paper
Figure 14. Real-world visualizations of Folding Box for the four WAM interfaces. Selected frames
Figure 15 of the paper
Figure 15. Real-world visualizations of Tabletop Cleaning for the four WAM interfaces. The se
Figure 16 of the paper
Figure 16. Case study of critical operations in Folding Clothes, Folding Box, and Tabletop Cleaning.
Figure 17 of the paper
Figure 17. Action-conditioned future-world prediction on Folding Clothes. UniW(A)M in AC-WM
Figure 18 of the paper
Figure 18. Action-conditioned future-world prediction on Folding Box. The comparison emphasizes
Figure 19 of the paper
Figure 19. Action-conditioned future-world prediction on Tabletop Cleaning. UniW(A)M retains

Figures are the authors', reproduced from the paper.

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