MV-WAM: Manifold-Aware World Action Model with Value Augmentation
1. Paper overview
In one sentence: MV-WAM improves visually shifted manipulation by giving coupled video and action experts different prediction targets, while its value-guided recovery remains operationally underspecified. e-probleme-architecturee-objectivese-valuee-protocole-simulatione-ablations
| At a glance | What to know |
|---|---|
| Research problem | Author claim Video priors improve clean-scene manipulation more reliably than action robustness under visual shifts. The authors attribute this imbalance to heterogeneous visual and action geometry, and seek to transfer action-free video experience into executable, language-conditioned action chunks. e-problem |
| Core mechanism | Source description A layerwise Mixture-of-Transformers joins distinct Video and Action-Value Experts through shared attention with directional modality access. e-architecture |
| A key reported result | RoboTwin 2.0: 50-task Clean/Random manipulation: 84.0 Clean; 55.7 Random Mean task success rate (%). Aloha AgileX; 100 rollouts per task per condition. WAMs receive Random video pretraining but only Clean action supervision; VLA baselines use 50 Clean demonstrations per task. HALO: 80.5/26.4; BagelVLA: 75.3/20.5. The HALO advantages are 3.5 and 29.3 percentage points. This tests transfer without randomized action labels, not without randomized visual exposure. Table 7 shows losses on individual tasks despite the best average. e-protocole-simulation |
| Reading caution | Source description Larger model scales are untested; sparse rewards may make Monte Carlo values noisy. Reported failures include wrong-arm choice for Adjust Bottle and inadequate contact precision for Move Can Pot. e-limitationse-failures |
Core contributions
- Source description
A layerwise Mixture-of-Transformers joins distinct Video and Action-Value Experts through shared attention with directional modality access. e-architecture
- Source description
Separate velocity and clean-endpoint losses, plus Monte Carlo value supervision and inference-time rollback, address optimization and execution respectively. e-objectivese-value
- Reader analysis
The loss-target ablation supports the chosen parameterization. It does not establish that manifold curvature is the unique causal explanation for improved robustness. e-ablationse-geometrye-theory
Figure 2. Two expert streams share context through an asymmetric mask and retain different prediction targets. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the two vertical stacks in panel (a): the VAE supplies video latents on the left, while state, noisy action and noisy value tokens enter the right. Snowflakes mark frozen VAE, umT5 and SigLIP2 modules; flames mark trained components. Read the small attention matrix by query row and context column. Blue video rows access the visual block, yellow action rows access video and actions, and the green value row c accesses every shown block. The filled blocks agree with Section 3.1's modality hierarchy. The right-hand panels show video pretraining followed by joint training; the bottom panel sketches a value drop, rollback and resampling. e-architecturee-objectivese-value
What it supports. The architecture supplies a concrete route for video experience to affect action generation inside shared transformer layers. Distinct expert parameters coexist with direct layerwise interaction. This supports an integrated MoT reading of the model and explains why counting two experts alone is insufficient to establish a separate world-model/policy pipeline.
Where the evidence stops. The drawn mask is blockwise, without a triangular time mask inside the video block; it does not demonstrate one-to-one frame causality. The rollback sketch also leaves physical scene restoration unspecified.
2. Motivation
2.1 The problem and the proposed response
Video priors improve clean-scene manipulation more reliably than action robustness under visual shifts. The authors attribute this imbalance to heterogeneous visual and action geometry, and seek to transfer action-free video experience into executable, language-conditioned action chunks. e-problem
2.2 What this reading follows
A robot can imagine a plausible future and still issue the wrong movement. MV-WAM addresses this gap by coupling a video expert to an action-value expert, letting actions read predicted visual context while protecting the video stream from action inputs. Its distinctive training choice is to predict video velocity but clean action endpoints. A learned progress estimate then monitors execution and can trigger rollback. The strongest evidence comes from randomized RoboTwin scenes with no randomized action supervision, although randomized videos are available during pretraining. Read the results alongside the ablations: prediction targets explain a much larger measured difference than adding value regulation. e-probleme-architecturee-objectivese-valuee-protocole-simulatione-ablations
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | IDM |
| Quadrant | Q4 · Dual-system × IDM |
3.1 Evidence-based assessment
Conflicts with the recorded classification
WAM is supported. Layerwise shared self-attention and coupled denoising support one integrated MoT model despite distinct expert parameters, conflicting with a standalone Dual-system interpretation. Actions have inverse-dynamics-like access to predicted futures, but no separate post-video IDM stage is specified. Monte Carlo value supervision alone does not establish RL policy post-training. The catalog snapshot is preserved for review. e-architecturee-objectivese-valuee-training
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Follow visual context into the action stream
Start with the distinction between prediction and execution. The video expert produces future visual content; the action expert produces a chunk of robot controls, which is then executed before new observations arrive. Section 3.1 connects the two through shared attention inside every paired transformer layer. The mask gives action tokens access to the video stream while preventing video queries from reading action tokens. My reading is that this provides inverse-dynamics-like conditioning within one integrated generator: actions are grounded in an imagined future, but a separate policy is not invoked after a completed video rollout. The value token occupies a third information position because it can inspect both streams. This architecture, rather than joint training alone, motivates questioning the catalog's Dual-system label. e-controle-architecturee-objectivese-value
5.2 Read the manifold claim through the actual losses
The operational change is easy to locate in Eqs. (2)–(3): video predicts a velocity target, while the action network predicts the clean action endpoint directly. Appendix F adds a crucial qualification. Computing endpoint error in velocity space introduces a 1/t² weighting, which emphasizes noise times near the data endpoint; MV-WAM removes that weighting from action supervision. The authors motivate this choice through different visual and action geometry. My interpretation is that the experiment supports a useful optimization choice more directly than a necessity theorem: changing targets also changes effective loss scaling, and the measured geodesic-distance ratio is explicitly not sectional curvature. A reproduction should therefore measure gradient balance across noise times as well as final success before attributing the entire improvement to intrinsic geometry. e-objectivese-samplere-geometrye-theorye-ablations
5.3 Turn progress estimates into a testable recovery rule
The value target is a discounted sum of trajectory rewards, using γ = 0.99. During execution, the policy remembers the largest earlier predicted value and compares the current estimate with that peak. Section 3.2 triggers rollback when the drop exceeds δ, then restores the cached checkpoint and samples a new action chunk. The default δ is 0.20. This creates two potential benefits: value prediction may improve representations during training, and rollback may change decisions during execution. The published no-value comparison combines these effects. My reading is that neither can be assigned the entire gain without an additional control. Physical recovery also needs an explicit meaning for restoring a checkpoint after objects have moved; the supplied description gives no actuator-level or environment-restoration procedure. e-valuee-ablationse-hardware
5.4 Training and inference
During training
The 1.9B parameter count excludes umT5 and SigLIP2. Video pretraining uses 20,000 steps, batch 1,024, AdamW, learning rate 10⁻⁴, weight decay 10⁻² and 240 × 320 video. Joint fine-tuning uses 15,000 steps, batch 512, learning rate 10⁻⁴ and action chunks of 32; the rate stays constant for 20% of steps, then decays linearly. e-training
The general pretraining corpus totals approximately 7,500 hours: EgoDex, Agibot, RoboMind, RoboCoin, Open X-Embodiment and 200 hours of private data. RoboTwin additionally specifies video pretraining from 50 Clean plus 500 Random demonstrations per task, followed by Clean-only paired action supervision. e-datae-protocol
Value tokens regress discounted trajectory returns with γ = 0.99. The paper does not specify a separate reinforcement-learning policy-improvement objective or fully detail reward construction and value-loss normalization. e-valuee-objectivese-training
During inference
Joint denoising defaults to five steps. Three steps attain 81.2%/54.2% Clean/Random versus 84.0%/55.7% at five. Reported throughput is 17.9 actions/second on one A800; this is not a measured rollback-inclusive deployment latency. e-ablationse-training
After executing action chunks, observations close the control loop. A rolling state-action-value buffer tracks peak progress. A value drop beyond δ triggers restoration of the peak checkpoint and stochastic resampling; default δ is 0.20. How a changed physical scene is restored is unspecified. e-controle-valuee-ablationse-hardware
5.5 Implementation flow
- Encode observations and instructions
WoW-1.3B initializes the Video Expert; a spatiotemporal VAE supplies visual tokens. umT5 conditions both experts. The Action-Value Expert alternates language cross-attention in odd layers and SigLIP2 visual conditioning in even layers. Figure 2 marks the VAE, umT5 and SigLIP2 frozen. e-architecture
- Align and route modalities
Both experts have 30 blocks. Action-side hidden width 768 projects to video width 1536 inside shared attention; its RoPE scale is one quarter of the video's. Video queries read video only; action/state queries read video and action context; value queries read the combined context. e-architecturee-training
- Match prediction targets
Video predicts the flow velocity, while actions directly regress clean endpoints. Appendix F removes the endpoint loss's velocity-space 1/t² weighting to avoid action-loss domination near the data endpoint. This changes supervision, while endpoint-induced velocity remains part of sampling. e-objectivese-sampler
6. Experiments & results
MV-WAM couples video, action and value generation through asymmetric shared attention. It trains video on velocity targets and actions on clean endpoints, then uses predicted progress to trigger rollback. Its strongest evidence is improved RoboTwin Random success; geometric necessity and reliable physical rollback remain less established (e-architecture, e-objectives, e-value, e-simulation, e-theory).
6.1 Read the original evidence
Table 7. The full task table exposes both MV-WAM's best average and substantial task-level weaknesses. Original paper, p. 18 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each task row contains a Clean/Random pair for every method; keep those two conditions separate when scanning. Begin at the Average row and compare MV-WAM's final pair with HALO's pair, then move upward to inspect individual tasks. Section 4.2 specifies 100 rollouts per task and condition. WAM video pretraining sees both Clean and Random demonstrations with action labels unused in that stage, whereas paired action fine-tuning uses Clean demonstrations. Thus the Random column measures transfer of visually varied experience into actions without randomized action supervision. It should not be read as evaluation with no prior randomized visual exposure. e-protocole-simulation
What it supports. MV-WAM reports 84.0% Clean and 55.7% Random, compared with HALO's 80.5% and 26.4%. The Random advantage is 29.3 percentage points. Variation remains large: MV-WAM reaches 88% Random on Blocks Ranking RGB, but only 2% on Open Microwave and 11% on Move Stapler Pad.
Where the evidence stops. The main text's claim of superiority across all tasks exceeds the table: several individual cells favor baselines. The table establishes the strongest reported average under this protocol, without repeated-seed uncertainty estimates.
Table 3. Physical execution improves overall, while folding succeeds in only two of ten trials. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each column as a different physical task on the TienKung robot. The first task is shortened to Pick Backbag in the table and called Pick Backbag & Coffee in Section 4.4. The table caption specifies ten trials per task, making each ten-percentage-point increment one successful trial. Compare methods within a task before reading Mean SR: Drop Cloth and Pick Cloth are easy for MV-WAM in this small sample, whereas Fold Cloth remains difficult for every model. The experiment uses three-view observations and shared fine-tuning demonstration sets across methods, with varying object configurations during evaluation. e-real-protocole-real-resultse-hardware
What it supports. MV-WAM's task success rates are 90%, 100%, 100% and 20%, averaging 77.5%; π₀ and RDT average 42.5% and 32.5%. These are physical execution results. The folding result is two successes in ten, compared with one for π₀ and none for RDT, which sharply limits claims about reliable deformable manipulation.
Where the evidence stops. Section 4.4 reports 100 demonstrations per task, while Appendix C reports 300. The supplied revision does not resolve that conflict. Ten evaluation trials per task provide limited evidence about deployment variability.
Figure 9. Selected failures separate a wrong imagined action from insufficient execution precision. Original paper, p. 20 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow each strip from left to right. For Adjust Bottle, the upper strip is explicitly labeled Predicted videos and the middle strip execution progress; the red connector highlights their relationship. The wrong-hand annotation belongs to the source graphic, and the accompanying text specifies an intended left-hand action but a selected right hand. The bottom strip contains execution frames for Move Can Pot and marks weak precision. Unlike the first example, it does not display a separate predicted-video strip. These layouts make it possible to distinguish the authors' semantic-planning diagnosis from their low-level contact-accuracy diagnosis without treating every image as an imagined future. e-failurese-value
What it supports. The reported failures show two remaining boundaries: a plausible predicted motion can select the wrong manipulator, and a generally correct motion direction can lack the precision needed for contact. Better average Random success therefore leaves both embodiment-aware reasoning and fine-grained action control as unresolved capabilities.
Where the evidence stops. These are selected illustrations, with no frequency estimate or matched successful control. They do not show value traces, rollback triggers or recovery attempts, so the images cannot establish why value regulation failed to prevent either outcome.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboTwin 2.0: 50-task Clean/Random manipulation Aloha AgileX; 100 rollouts per task per condition. WAMs receive Random video pretraining but only Clean action supervision; VLA baselines use 50 Clean demonstrations per task. | 84.0 Clean; 55.7 Random Mean task success rate (%) | HALO: 80.5/26.4; BagelVLA: 75.3/20.5. The HALO advantages are 3.5 and 29.3 percentage points. This tests transfer without randomized action labels, not without randomized visual exposure. Table 7 shows losses on individual tasks despite the best average. e-protocole-simulation |
| RoboTwin 2.0: transfer to 10 held-out tasks Train on 40 tasks; the paper specifies 22,000 demonstrations. Zero-shot excludes held-out task demonstrations during fine-tuning; few-shot adds demonstrations per test task. | 0-shot: 55.6/54.0; 10-shot: 77.1/74.6 Mean task success rate, Clean/Random (%) | 5-shot: 76.2/75.7. Adaptation helps the average, but Random success is not monotonic in shot count. This differs from the 50-task protocol. e-transfer |
| Prediction-target and value ablations Same training setting and 15,000 steps for target comparisons; RoboTwin Clean/Random. | Full: 84.0/55.7; without value: 83.4/54.5 Success rate, Clean/Random (%) | Noise-pred: 15.0/15.0; flow-pred: 9.0/6.0. Target assignment has the larger observed effect. Value adds 0.6/1.2 percentage points, but the comparison does not separate value supervision from rollback; no seed uncertainty is provided. e-ablations |
| TienKung: four physical manipulation tasks Identical demonstration sets across methods; 10 trials per task with varying object configurations. Demonstration count conflicts between main text and appendix. | Pick Backbag: 90; Drop Cloth: 100; Pick Cloth: 100; Fold Cloth: 20; mean: 77.5 Task and mean success rate (%) | π₀ mean 42.5; RDT mean 32.5. Fold Cloth: 10 and 0 respectively. These are executed robot outcomes. Folding remains weak: 20% represents two successes in ten, not robust deformable manipulation. The text calls the first task Pick Backbag & Coffee. e-real-protocole-real-results |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 3. Target choice dominates the reported ablation differences; value regulation provides a smaller increment. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the legend consistently: pale bars are Clean and dark bars Random. Panel (a) places noise prediction, flow prediction, the model without value, and the full model on one axis. Section 4.3 states that the target comparisons use the same training setting and 15,000 steps. Panel (b) varies shared denoising steps for both experts, and panel (c) varies the rollback threshold δ. Read printed values rather than comparing bar heights across panels: the second and third axes begin above zero. The default configuration uses five denoising steps and δ = 0.20; the largest displayed threshold is a sensitivity test, not the reported default. e-ablationse-objectivese-value
What it supports. Full-model Random success is 55.7%, versus 15.0% for noise prediction and 6.0% for flow prediction. Removing value leaves 54.5%, a much smaller difference. Three denoising steps give 54.2% Random versus 55.7% at five, indicating a useful performance/step-count tradeoff within the tested setup.
Where the evidence stops. No error bars establish the reliability of the small value gains. Removing value changes supervision and the available execution feedback together, so this comparison does not isolate the benefit of rollback itself.
Figure 7. The representation diagnostic motivates the geometry hypothesis without proving its causal explanation. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare color mixing within each panel; the t-SNE coordinates have no physical units. Blue filled points denote Clean conditions and orange outlined points Random conditions. The authors extract shared-attention features, average tokens per frame, and subtract each task's mean feature before projection. They sample three episodes and five frames per episode for each of 50 tasks, yielding roughly 750 points per condition, and use perplexity 50. The left panel's colors overlap considerably; the right panel separates them. This is a representation diagnostic. The separate distance-ratio analysis and Table 5 quantify the paper's geometry proxy using a different procedure. e-geometrye-theorye-ablations
What it supports. The visualization is consistent with action representations changing more than video representations under the tested condition shift. It supplies a motivation for modality-specific training objectives. Its strongest supported conclusion concerns the displayed representation separation; action-execution improvement must be assessed through the benchmark and ablation results.
Where the evidence stops. The source does not clearly identify the diagnostic checkpoint. A two-dimensional projection cannot establish causal brittleness or sectional curvature; Appendix D.2 explicitly acknowledges that the separate empirical geometry proxy is not a sectional-curvature estimate.
7. Analysis & limitations
7.1 What the evidence leaves open
Larger model scales are untested; sparse rewards may make Monte Carlo values noisy. Reported failures include wrong-arm choice for Adjust Bottle and inadequate contact precision for Move Can Pot. e-limitationse-failures
The reported geometry statistic is a geodesic/Euclidean distance proxy: 3.8 ± 0.6 for vision versus 1.3 ± 0.2 for actions, p < 0.001 across tasks. Appendix D.2 concedes it is not sectional curvature. The theoretical argument assumes a curvature-dependent penalty and cross-modal regularity coupling; the measurements do not prove those assumptions. e-geometrye-theory
Figure 7 demonstrates a projection-level condition separation, not causal proof of policy failure. Success tables and ablations lack repeated-seed uncertainty, and no dedicated causal-mask ablation is reported. e-geometrye-ablationse-simulatione-real-results
7.2 Questions for discussion
- Would gradient-scale-matched objectives retain the large prediction-target advantage?
- Does value supervision help when rollback is disabled, and can rollback recover physical scenes without external restoration?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires the video backbone, frozen encoders, dataset mixture, paired demonstrations and modality-specific losses. The private corpus portion, unspecified dataset versions/mixing ratios, training GPU budget, software versions and seeds limit exact replication. e-architecturee-traininge-data
Resolve 100 demonstrations/task in Section 4.4 versus 300 in Appendix C. The physical platform uses three Orbbec Gemini 330 cameras and 16 active control dimensions, while Section 2.2 standardizes actions to 128 dimensions; the exact packing is not detailed. e-real-protocole-hardwaree-control
Clarify the solver before replication: Table 6 gives x₁ minus the predicted clean endpoint for x₀ inference velocity, but adjacent text refers to denominator clipping and Eq. (7) uses an endpoint velocity divided by t. Eq. (2)'s explanatory interpolation also mixes t and τ. Physical checkpoint restoration and reward labeling need explicit operational definitions. e-samplere-objectivese-valuee-hardware
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Control loss scaling while changing the action target
Reader-proposed check: keep the pretrained video checkpoint, data split, attention mask, optimizer, training steps and five-step evaluation fixed. Compare direct endpoint supervision with velocity supervision and a velocity-loss variant matched for action-gradient magnitude across noise-time bins. Repeat with multiple seeds; log branch gradients, action error and Clean/Random success. Clarify Table 6's solver convention before fixing the common sampling implementation. If gradient matching closes the success gap, the result would support optimization balance as an alternative explanation; a persistent gap would strengthen the case for the target choice. e-objectivese-samplere-traininge-protocole-ablations
Check 2: Separate value learning, resampling and state restoration
Reader-proposed check: compare a no-value policy, a value-trained policy with rollback disabled, and the same value-trained checkpoint with δ = 0.20 rollback enabled, using matched scene seeds and action/resampling budgets. Record success, false triggers, recovery rate, extra actions and latency. In simulation, explicitly distinguish exact checkpoint restoration from re-observing the current scene and resampling without restoring objects. If the benefit appears only with exact restoration, it supports reset-assisted recovery under that environment capability. Improvement without restoration would provide stronger evidence for an executable recovery mechanism. e-valuee-ablationse-protocole-hardware
8.3 Reading coverage
Visual audit: The title/author page and all pages supporting retained method, training, numerical, evaluation, hardware and reproduction claims were rendered and visually inspected. This included Figures 1–9 and Tables 1–7; six final original crops were individually inspected. Figure 2's mask was cross-checked against Section 3.1, and rollback against Section 3.2. Table 6 and Eq. (7) retain an unresolved solver-description discrepancy. Reference-only pages 11–13 were read in the text chunks, without image inspection. Separate supplements and code were outside the supplied reading material.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 14, 15, 16, 17, 18, 19, 20. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract
- 1 Introduction
- 2 Problem Definition and Analysis (2.1–2.3)
- 3 Method (3.1–3.2)
- 4 Experiments (4.1–4.4)
- 5 Conclusion and Limitations
- References
- A Related Work
- B Pre-training Dataset Details
- C Real-World Experiment Setup (C.1)
- D Generalization Analysis in Unified World Action Models (D.1–D.2)
- E Algorithm
- F Discussion
- G Simulation Results
- H Broader Impact
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Identity: the inspected title page identifies arXiv:2606.21088v1 [cs.RO], 19 June 2026. The title and all 16 authors agree with the catalog after normalizing name order and typography. No different revision or edition was supplied or compared.
- All seven supplied text chunks were read individually, covering all 20 PDF pages and appendices. Extraction does not reconstruct figures; the retained PDF was therefore visually inspected separately. Reference-only pages 11–13 were read as text, without a page-image pass.
- Separate supplemental material availability has not been fully verified; no separate supplements were supplied.
- No code, external links or datasets were inspected, and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title, author block, affiliation footnote and arXiv margin
The exact title, 16 authors, two institutions and arXiv:2606.21088v1 dated 19 June 2026 are printed on the title page.
Go to primary source ↓e-problemPDF pp. 2–4, Sections 1, 2.1 and 2.3
The authors motivate visual-shift generalization and attribute impaired action robustness to heterogeneous modality manifolds.
Go to primary source ↓e-controlPDF pp. 3–4, Sections 2.1–2.2 and Eq. (1)
The formulation uses observation history, language and executed action chunks with feedback; actions are standardized into 128 dimensions and flow matching interpolates clean targets with noise.
Go to primary source ↓e-architecturePDF pp. 4–5, Section 3.1; Figure 2(a), mask matrix and frozen/trained legend
Two modality-specific DiT experts interact through shared attention. The visible mask allows video-to-video, action queries over video/actions and value queries over all streams. Encoders, RoPE scaling and frozen modules are described.
Go to primary source ↓e-objectivesPDF p. 5, Section 3.1, Eqs. (2)–(3); pp. 17–18, Appendix F, Eq. (7)
Video uses velocity regression; actions use unweighted clean-endpoint regression, with both modality weights one. Appendix F explains removing 1/t² endpoint-loss weighting; p. 5 mixes t and τ in its interpolation text.
Go to primary source ↓e-valuePDF p. 6, Section 3.2, Eq. (4) and Value-Guided Rollback paragraph
Value targets are Monte Carlo returns with discount 0.99. A drop below peak minus δ triggers cached-state rollback and resampling; physical state restoration and detailed reward construction are not specified.
Go to primary source ↓e-trainingPDF pp. 6–7, Section 4.1
Reports 1.9B parameters excluding encoders; 30-block experts, pretraining/fine-tuning schedules, dimensions, chunk size and 17.9 actions/s on a single A800. Training hardware budget, seeds and software versions are not listed.
Go to primary source ↓e-dataPDF p. 14, Appendix B, Table 4
Estimated data hours are EgoDex 800, Agibot 2,500, RoboMind 1,000, RoboCoin 1,000, Open X-Embodiment 2,000 and private data 200, totaling 7,500; mixture ratios and dataset versions are absent.
Go to primary source ↓e-protocolPDF p. 7, Section 4.2, Benchmark and Protocol
WAM video pretraining uses 50 Clean plus 500 Random demonstrations/task; paired fine-tuning is Clean-only. VLA baselines use 50 Clean demonstrations/task. Evaluation uses 100 rollouts/task/condition.
Go to primary source ↓e-simulationPDF p. 7, Table 1 and results; p. 18, Table 7, Average, Blocks Ranking RGB, Open Microwave and Move Stapler Pad rows
MV-WAM averages 84.0/55.7 Clean/Random; HALO 80.5/26.4 and BagelVLA 75.3/20.5. MV-WAM Random is 88 for Blocks Ranking RGB, 2 for Open Microwave and 11 for Move Stapler Pad; it is not best in every cell.
Go to primary source ↓e-transferPDF p. 8, Table 2, Mean SR row and Section 4.2 continuation
The 40/10 task split uses 22,000 training demonstrations. Mean Clean/Random success is 55.6/54.0 at zero-shot, 76.2/75.7 at five-shot and 77.1/74.6 at ten-shot.
Go to primary source ↓e-ablationsPDF p. 8, Section 4.3; p. 9, Figure 3(a–c)
Figure 3 prints target/no-value/full results, denoising steps and thresholds. At δ=0.20 full success is 84.0/55.7 versus 83.4/54.5 without value; three steps yield 81.2/54.2. No error bars or isolated mask ablation are supplied.
Go to primary source ↓e-real-protocolPDF p. 9, Section 4.4 and Figure 4; p. 15, Appendix C.1
The main text specifies 100 demonstrations/task at 30 FPS and 10 evaluation trials/task; Appendix C specifies 300 demonstrations/task at the same rate. Both describe shared fine-tuning data across methods.
Go to primary source ↓e-real-resultsPDF p. 10, Table 3, all rows and columns; p. 9, Section 4.4
MV-WAM success percentages are 90, 100, 100 and 20 across Pick Backbag, Drop Cloth, Pick Cloth and Fold Cloth; mean 77.5 versus π₀ 42.5 and RDT 32.5. Each task has ten trials.
Go to primary source ↓e-hardwarePDF p. 15, Appendix C.1, Figure 5 and Eq. (5)
TienKung uses three Orbbec Gemini 330 RGB-D cameras, with two wrist views and one stationary global view. RGB conditions the model; the active action vector has seven joint deltas and one gripper command per arm.
Go to primary source ↓e-geometryPDF p. 16, Appendix D.1, Figure 7 and Table 5
Features from shared attention are token-mean pooled and task-centered: 50 tasks, three episodes and five frames per condition, about 750 points/condition, t-SNE perplexity 50. Geometry uses PCA-50 for vision, a 10-neighbor geodesic graph and native actions; proxy mean/std are 3.8/0.6 and 1.3/0.2, with Mann–Whitney p<0.001.
Go to primary source ↓e-theoryPDF p. 4, Section 2.3; p. 17, Appendix D.2
The authors invoke a curvature-dependent generalization penalty for Lipschitz functions on compact Riemannian manifolds. Appendix D.2 explicitly distinguishes the empirical distance-ratio proxy from sectional curvature.
Go to primary source ↓e-samplerPDF p. 17, Appendix E, Eq. (6), Table 6, clipping sentence and Appendix F Eq. (7); p. 18, Appendix F continuation
Table 6's x₀ inference velocity is x₁−predicted x₀, while surrounding discussion mentions denominator clipping and Eq. (7) uses division by t. Appendix F states endpoint-induced velocity is still used during sampling.
Go to primary source ↓e-limitationsPDF p. 10, Section 5, Limitations paragraph
The authors identify untested larger scales and noisy Monte Carlo values under sparse rewards as limitations.
Go to primary source ↓e-failuresPDF pp. 19–20, Appendix G, Failure analysis and Figure 9
Adjust Bottle illustrates a predicted wrong-arm choice: the authors specify left as intended and right as selected. Move Can Pot illustrates insufficient contact precision. These are selected qualitative examples, not failure-frequency estimates.
Go to primary source ↓8.5 Primary sources
MV-WAM: Manifold-Aware World Action Model with Value Augmentation ↗
PDF · 10,463 extracted words
Source fingerprint
8360e2e4fabb4fed2c537f61da2d3ebb08759caed47869f63839f089bdc8b463