Selective Cross-View Consistency for World Action Models: Held-Out Viewpoint Robustness Without Test-Time Camera Information
1. Paper overview
In one sentence: Selective agreement across camera views improves simulated control beyond trained viewpoint ranges, while leaving future images view-dependent and offering no interpolation benefit. E02E03E05E07E08E12E13E16E17
| At a glance | What to know |
|---|---|
| Research problem | Source description A WAM predicts outputs with different transformation laws: changing the camera should change the future scene image but preserve base-frame actions and other physical targets. Uniform consistency therefore conflicts with correct video supervision. Evaluation must also distinguish benefits from camera exposure, the added consistency objective, and access to a stable wrist view. E02E03E08E10 |
| Core mechanism | |
| A key reported result | Held-out orbital extrapolation: First seed: SCVC 76.0%; improvement +12.20 percentage points, 95% CI [7.40,17.00]. Second seed: +15.50 points [11.70,19.40]. Closed-loop success and paired task-bootstrap difference. Scene-only LIBERO-Plus; |h| in (62,75] degrees; 100 tasks, ten trials per task; matched training arms. First-seed pair-trained control: 63.8%. Both seeds support extrapolation improvement attributable to the added objective under this controlled protocol. E07E08E12E13 |
| Reading caution | Source description All WAM execution evidence is simulated and scene-only. Consistency is trained on demonstration states, leaving failure/recovery states and wrist-camera pose perturbations untested. Real deployment also requires obtaining sufficiently matched cross-view training observations. E17 |
Core contributions
- Reader analysis
The masking audit shows that substantial camera-track success can persist without the perturbed scene-camera input, motivating scene-only mechanism experiments. E10
Figure 1a. Agreement belongs on physical targets that remain fixed when the camera moves. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with physical state s at the left: the two observations differ in camera, not in the underlying scene and robot state. Both enter Dθ with shared parameters θ, noise level σ, and noise realization n. The blue arrows compare the action chunk a₁:K, future proprioception q′, and value V through the invariant-frame selector Π_I. The crossed orange link is equally important: predicted future images retain separate camera-specific targets. Equation (1) confirms that both complete branches receive supervised denoising loss; only the extra agreement penalty is selective. This is the paired training construction. Deployment still receives the ordinary scene image, proprioception, and instruction. E02E03E20
What it supports. A single denoiser can produce outputs with different transformation laws. SCVC exploits that structure by constraining the action-related frames while allowing future images to move with the camera. The two depicted branches share a model; their presence does not make this a two-model control architecture.
Where the evidence stops. Action invariance depends on base-frame commands, matched physical states, task-informative views, and unchanged semantics. Camera settings are used to render and partition data; the policy does not consume them as conditioning inputs.
2. Motivation
2.1 The problem and the proposed response
A WAM predicts outputs with different transformation laws: changing the camera should change the future scene image but preserve base-frame actions and other physical targets. Uniform consistency therefore conflicts with correct video supervision. Evaluation must also distinguish benefits from camera exposure, the added consistency objective, and access to a stable wrist view. E02E03E08E10
2.2 What this reading follows
Moving a camera changes what a robot sees without necessarily changing what it should do. A world action model complicates that distinction because it predicts both future images and actions. SCVC makes the distinction explicit: same-state views should agree on base-frame actions, future proprioception, and value, but should keep their own future images. Shared diffusion noise makes the comparison meaningful. The paper couples this objective with a carefully carved camera distribution and an identically pair-trained control. Its strongest result concerns extrapolation beyond trained viewpoints. Interpolation, task-level regressions, and dependence on simulated training pairs set substantial limits on the claim. E02E03E05E07E08E12E13E16E17
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 | One Model |
| Prediction paradigm | Joint prediction |
| Quadrant | Q1 · One Model × Joint prediction |
3.1 Evidence-based assessment
Supports the recorded classification
One denoiser directly predicts actions, proprioception, future images, and value in a shared latent sequence, supporting One Model × Joint prediction. Paired branches are a training construction, not separate models or inverse dynamics. Multiview supervision supports the catalog subcategory, but the paper has no explicit 3D reconstruction or camera-conditioned deployment. E02E03E07
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 First establish which outputs a camera change should preserve
A same-state pair holds the robot and scene fixed while changing the rendering camera. That pairing alone does not make every label invariant. The paper assumes delta end-effector commands are expressed in the robot base frame, with both views informative and task semantics unchanged. Under those conditions, the 16-step action chunk, future proprioception, and value can share labels, while the future scene cannot. Cosmos encodes these targets in one seven-frame latent sequence alongside blank and conditioning frames. The two training branches therefore use identical physical targets but distinct current and future scene images. At deployment, the scene image, proprioception, and language condition direct joint prediction. The paper leaves the action execution cadence and inference-time role of value unspecified; the diagram does not establish a separate planning loop. E02E03E09E20
Figure 1b. The location of a held-out camera determines which generalization claim is possible. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the blue ranges as training coverage and the unfilled interior gaps as interpolation tests. Azimuth keeps absolute h from 0 to 62 degrees except the open interval (20,30); views above 62 through 75 are extrapolation. Dolly training spans 100–175 except (128,142), with extrapolation above 175 through 200. Endpoint and elevation are sparse rather than continuous: blue dots indicate off-benchmark training values, while held-out official settings use open circles or orange squares. The text specifies endpoint training at ±3, ±5, and ±7, and elevation training at 8. Elevation 15 therefore supplies one extrapolation setting, not a continuous sweep. E07E08E20E24
What it supports. Camera diversity and camera generalization are different experimental conditions. The carved manifest lets an identically exposed control and SCVC be compared inside trained regions, inside withheld holes, and outside the training envelope. Any benefit over that control tests the objective beyond the contribution of pair exposure itself.
Where the evidence stops. This is not nominal-training zero-shot evaluation: both comparison arms see perturbed viewpoints. The schematic also omits compound buckets and the small ±2 endpoint bucket discussed in the text; it should not be read as the complete task inventory.
5.2 Then ask what the extra agreement term actually penalizes
Lemma 1 separates an invariant prediction pair into its mean and its cross-view residual. Supervision already penalizes the residual; consistency increases that coefficient from one to 1 + 4λ. Thus the algebra is more precise than the accompanying statement that supervision constrains only the mean. For a future-scene target that legitimately differs across cameras, the same extra pressure becomes bias: Proposition 1 shrinks the optimal difference by 1/(1 + 4λ). As a reader calculation, λ = 2 leaves one ninth of that difference in the pointwise model, not a guaranteed ratio for a large trained WAM. Noise must also be paired. Proposition 2 shows that independent noise penalizes legitimate noise dependence at view invariance, while explicitly withholding a globally positive-sign claim away from that regime. E03E04E05E06
5.3 Finally separate added consistency from camera exposure
The correct comparator already sees the same perturbed pairs and receives supervision on both branches. Starting both arms from the same nominal scene-only reference and using the same budget makes the consistency weight the experimental difference. Carving interior holes and exterior ranges then separates interpolation from extrapolation instead of calling every high camera score generalization. Orbital extrapolation improves in both training seeds; interpolation is null in the first and significantly worse in the second. A trained azimuth region also improves, so a reader should consider severity and performance headroom alongside camera novelty when interpreting the mechanism. The large positive spatial-task effect and negative object-task effects further argue for suite-level reporting. Aggregate preservation does not establish that every task, camera axis, or manipulation failure mode benefits. E07E08E09E12E13E15E16E17
5.4 Training and inference
During training
Consistency ramps linearly to λCV = 2.0 during the first 2% of training and applies only to demonstration samples. Rollout-mode samples remain single-view. The integrated manifest contains 583,648 same-state pairs. E03E07
A nominal scene-only reference is trained from the released Cosmos Policy LIBERO checkpoint, frozen, then used to initialize both comparison arms. Each receives 10,000 steps, effective batch 720, 7.2 million presentations, learning rate 5 × 10⁻⁵, and EMA; evaluation uses the final checkpoint. E08E09
During inference
SCVC leaves the deployment interface unchanged: one scene observation, proprioception, and instruction condition the joint denoiser. Pairing exists only during training; no camera metadata, depth, or additional view-generation procedure is required by the policy. E02E03
Actions are directly predicted within the joint target, and the reported success rates come from closed-loop simulator evaluation. The paper does not specify the executed fraction of each chunk, denoising-step count, or inference-time use of the value output. E02E09
5.5 Implementation flow
- Establish action-equivalent pairs
Render nominal and perturbed observations from the same physical state. Label sharing assumes base-frame actions, task-informative views, and unchanged task semantics. The authors report checking the controller convention and asserting bitwise-identical invariant targets. E02
- Jointly encode prediction targets
The scene-only Cosmos sequence contains seven latent frames: blank, current proprioception, current scene conditioning, actions, future proprioception, future scene, and value. Both pair branches share one denoiser. E02E20
- Apply selective consistency
Share the noise level and realization across branches. Supervise both complete predictions, but penalize their difference only on action, future-proprioception, and value frames. Future-scene predictions remain separately supervised. E03E05
6. Experiments & results
SCVC trains a joint video-action denoiser to agree across camera views on actions, future proprioception, and value, while preserving camera-specific future images. Matched simulated experiments support extrapolation gains beyond trained camera ranges; interpolation does not improve and regresses significantly in the second seed. Wrong-coordinate and wrist-masking diagnostics explain the objective and its evaluation contract.
The source provides a short 2,000-step wrong-coordinate image diagnostic and controlled-scale shrinkage measurements, but no converged full-scale shrinkage measurement or direct shared-versus-independent-noise closed-loop ablation. The full second-seed twelve-bucket table is not supplied. Shuffled-pair and real-robot corroboration is attributed to a separate flow-VLA study, not an experiment reproduced in this WAM paper. E05E06E17E19E24
6.1 Read the original evidence
Figure 4. Plausible predicted frames can coexist with unsuccessful closed-loop actions. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The horizontal axis is relative excess-FID, with larger values indicating worse predicted-frame fidelity; the vertical axis is action success. The metric subtracts a camera-conditioned oracle floor computed from ground-truth replays, and the dashed line marks zero relative excess-FID. Green squares trace orbital viewpoints, blue circles dolly changes, and orange triangles endpoint reorientation. Each marker represents one severity level. The orbital series is nearly vertical near the floor: control varies strongly while this image-distribution measure changes little. Dolly instead shows broader rightward displacement with poor control. Read the severity trend from Section VI-C because the rotated level annotation is potentially misleading. E11E22
What it supports. The text reports orbital action success falling from 53% to 3% while relative excess-FID remains between −0.07 and +0.16. That dissociation motivates an action-side objective, but it concerns the frozen scene-only reference. It is not a before-and-after evaluation demonstrating that SCVC repaired those predicted frames.
Where the evidence stops. The rotated 'severity L1 → L5' annotation points upward on the page, whereas Section VI-C describes decreasing success with increasing severity; individual levels are not labeled. Use the prose for ordering. Distributional fidelity also does not certify semantic plan correctness.
Table II. The matched comparison supports extrapolation gains and preserves the null results. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. N counts tasks, with ten trials per task for the camera buckets. Control and SCVC columns report success percentages; Δ reports the paired task-bootstrap difference in percentage points with a 95% interval. The bold azimuth interpolation and extrapolation rows are the pre-registered primary endpoints. Read them together: the former interval crosses zero, whereas the latter is wholly positive. Dolly and elevation extrapolation are secondary corroborating endpoints; endpoint reorientation is null. The in-distribution azimuth row is a distribution-matched improvement, not an unseen-view result. Finally, the bottom row evaluates nominal competence over four LIBERO suites with 500 episodes each and supplies no interval. E08E09E12E13E14E15E16E24
What it supports. On 100 orbital extrapolation tasks, SCVC reaches 76.0% versus the control's 63.8%, a +12.20-point difference [7.40,17.00]. Both arms received the same carved pairs and budget. Dolly and elevation extrapolation also improve, while the displayed azimuth interpolation result does not establish a benefit.
Where the evidence stops. This table shows the first seed and nine single-axis buckets, not the complete twelve-bucket protocol. The second seed's text reports azimuth extrapolation +15.50 points but interpolation −4.29 [−7.56,−1.22]. Task-bootstrap intervals do not capture training-seed variation or suite-level regressions.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Held-out orbital extrapolation Scene-only LIBERO-Plus; |h| in (62,75] degrees; 100 tasks, ten trials per task; matched training arms. | First seed: SCVC 76.0%; improvement +12.20 percentage points, 95% CI [7.40,17.00]. Second seed: +15.50 points [11.70,19.40]. Closed-loop success and paired task-bootstrap difference | First-seed pair-trained control: 63.8%. Both seeds support extrapolation improvement attributable to the added objective under this controlled protocol. E07E08E12E13 |
| Held-out orbital interpolation Azimuth hole (20,30) degrees in absolute magnitude; 49 tasks, ten trials per task. | First seed: SCVC 85.1%, difference −1.22 points [−4.29,1.84]. Second seed: 85.7%, difference −4.29 points [−7.56,−1.22]. Closed-loop success and paired task-bootstrap difference | Controls: 86.3% and 90.0%, respectively. The first comparison is inconclusive; the second shows a regression. Describing interpolation as uniformly preserved overstates the evidence. E07E12E13 |
| Held-out dolly extrapolation Scale in (175,200]; 99 tasks, ten trials per task; secondary endpoint. | First seed: SCVC 71.2%, +4.24 points [0.71,7.88]. Second-seed difference: +4.44 points [0.61,8.49]. Closed-loop success and paired task-bootstrap difference | First-seed control: 67.0%. A smaller extrapolation benefit replicates across seeds. E07E13E14 |
| Held-out elevation extrapolation Official elevation 15, with training at the off-benchmark value 8; 386 tasks, ten trials per task. | First seed: SCVC 63.7%, +8.76 points [6.40,11.11]. Second-seed difference: +4.84 points [3.11,6.58]. Closed-loop success and paired task-bootstrap difference | First-seed control: 54.9%. Positive effects replicate, but this axis tests only one extrapolated elevation and has no reported interpolation counterpart. E07E13E14E17 |
| Nominal task competence Four LIBERO suites, 500 episodes per suite. | First seed: SCVC 91.7%, difference −0.60 points. Second-seed difference: −0.2 points. Closed-loop success | First-seed control: 92.3%. Aggregate nominal competence changes little; no confidence interval is supplied for this comparison. E09E13E15 |
| Wrong-coordinate mechanism check Synthetic shared-trunk denoiser trained to convergence; separate short 2B WAM diagnostic. | Synthetic ratios at λ = 0.1, 0.5, 2.0: 0.715, 0.337, 0.116. The wrong-coordinate 2B arm reaches a future-scene ratio of 0.18 after 2,000 steps. Normalized cross-view future-target difference | Analytic ratios: 0.714, 0.333, 0.111. Selective and zero-consistency 2B arms remain near 0.98. Controlled results closely follow the shrinkage prediction. The short full-model diagnostic supports collapse of view-specific content, but does not measure converged full-scale shrinkage or executed-action degradation. E04E06 |
| Wrist-camera dependence audit Released wrist-enabled Cosmos policy on camera-perturbed tasks; 1,200 rollouts per condition. | Scene blacked: 66.6%; scene replaced by noise: 69.7%; wrist blacked: 0.2%. Closed-loop success | Unmasked audit baseline: 80.5%. Success without the scene view establishes that many tasks can bypass the perturbed sensor. Wrist-mask failure alone cannot distinguish reliance from corruption sensitivity. Table I reports a separate 79.5% camera aggregate, which should not be substituted for this audit baseline. E10 |
| Video-fidelity versus control diagnosis Frozen scene-only reference across orbital severity levels on the camera track. | Action success falls from 53% to 3%, while relative excess-FID stays between −0.07 and +0.16. Action success and relative camera-conditioned excess-FID | The fidelity diagnostic subtracts an oracle floor from ground-truth replays under the same camera. Stable distributional image fidelity does not establish correct plans or successful control. The text supplies the severity ordering; Figure 4's upward-rotated severity annotation is ambiguous against that ordering. E11E22 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 2. Forcing future scenes to agree removes legitimate differences between views. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each row is a held-out pair. Read the first three columns as ground truth, wrong-coordinate prediction, and selective prediction for view A, then repeat that comparison for view B in the last three columns. The selective columns retain the different camera geometries; the wrong-coordinate columns show blurred, overlapping scene content. Section IV-C describes the intervention precisely: the wrong-coordinate arm additionally applies consistency to the future-scene block. Both illustrated arms use the 2B WAM after 2,000 training steps and are decoded under shared noise. Compare corresponding objects and the arm silhouette within a row, rather than treating the rows as a time sequence. E04E06E21
What it supports. The examples support the predicted failure mode of enforcing equality on camera-dependent outputs. Preserving separate future targets allows the model to retain viewpoint-specific content. This visual diagnostic complements the controlled shrinkage experiment; it does not itself measure a numerical shrinkage coefficient or a change in executed-action success.
Where the evidence stops. The caption says both arms track the near-nominal view A, yet its wrong-coordinate column also visibly blurs details. Preserve that qualification. These short-run examples cannot establish converged full-scale shrinkage or closed-loop harm from the wrong-coordinate variant.
Figure 3. A stable wrist view can sustain a camera-track score when the scene view is removed. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. These bars audit the released wrist-enabled Cosmos policy, not the scene-only SCVC comparison. Read the first bar as the unmasked reference, the next two as distinct corruptions of the scene camera, and the last as removal of the wrist image. The experiment runs 1,200 rollouts per condition, and the caption identifies the error bars as Wilson 95% confidence intervals. The benchmark perturbs the scene camera, while the wrist camera remains attached to the arm. Consequently, the scene-masking conditions ask whether the supposedly perturbed information is necessary for success. They provide the strongest direction of evidence in this audit. E10
What it supports. Success remains 66.6% with a black scene image and 69.7% with scene noise, versus 80.5% unmasked. Thus many camera-perturbed tasks can be solved without scene observations. This supports the authors' choice to remove the wrist view when measuring the scene-view invariance targeted by SCVC.
Where the evidence stops. The 0.2% wrist-mask score may reflect both reliance and corruption sensitivity. Table I reports a separate 79.5% released-policy camera aggregate; it is not this audit's 80.5% baseline. These masking results do not quantify SCVC's gain.
7. Analysis & limitations
7.1 What the evidence leaves open
All WAM execution evidence is simulated and scene-only. Consistency is trained on demonstration states, leaving failure/recovery states and wrist-camera pose perturbations untested. Real deployment also requires obtaining sufficiently matched cross-view training observations. E17
Aggregate gains hide substantial heterogeneity: the discussion reports +40.8 points for spatial tasks but −11.6 for object tasks within primary extrapolation, and −40.4 for object tasks under an extreme compound shift. The proposed grasp-precision explanation is an interpretation rather than an isolated causal result. E16
Two seeds provide limited training-variation evidence; task-bootstrap intervals do not capture it. Second-seed interpolation regression qualifies the conclusion's preservation wording. Endpoint effects are null, and elevation has only one extrapolation point, despite prose referring broadly to interpolation counterparts. E13E14E17E23
The pointwise theorem and synthetic convergence experiment do not establish exact shrinkage in the deployed 2B model. The paper supplies neither a converged full-scale wrong-coordinate control nor a direct shared-versus-independent-noise execution ablation. E05E06E17
7.2 Questions for discussion
- Does SCVC principally help unseen geometry, or difficult viewpoints with performance headroom? The trained azimuth bucket also improves. [E12, E15]
- How should consistency respond when occlusion makes one view less informative or physical-state matching is approximate? [E02, E17]
- Can task-conditioned consistency preserve the spatial-task gains while preventing object-task regressions under compound camera shifts? [E16]
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the Cosmos Policy LIBERO checkpoint, nominal scene-only initialization, recoverable demonstration states, and camera-specific future targets. Manifest guards must enforce the disclosed azimuth and scale holes, endpoint training values {±3,±5,±7}, elevation 8, and held-out official evaluation settings. E02E07E09
Hardware, wall-clock cost, EMA decay, detailed optimizer settings, frozen-module policy, denoising steps, and replanning cadence are unspecified. Code/tooling is promised without a repository. Shuffled-pair and real-robot corroboration comes from a separate flow-VLA work whose cited identifier remains 2608.XXXXX. E03E09E18E19
Proposed checks: compare selective versus wrong-coordinate losses under shared versus independent noise on controlled targets, then repeat matched simulator evaluation with additional training seeds and suite-level reporting. Neither check has been run here. E04E05E06E08E16E17
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: Factor the coordinate selector and noise pairing apart
Reader-proposed, not performed: recreate the controlled shared-trunk denoiser with one exactly invariant target block and one exactly camera-dependent block. Compare selective versus additional covariant consistency, each under shared versus independent noise, plus the identically pair-supervised λ = 0 control. Match architecture, target pairs, optimizer, convergence criterion, and training seeds; sweep λ over 0.1, 0.5, and 2.0. Measure held-out normalized view differences separately from invariant-target error. Wrong-coordinate shared-noise runs should approach the reported shrinkage trend, while independent noise should add residual pressure near invariance. Failure to separate those effects would challenge either the implementation or the proposed noise-matching explanation, without yet settling robot-control performance. E03E04E05E06
Check 2: Test whether extrapolation gains survive training variation and task heterogeneity
Reader-proposed, not performed: retrain matched λCV = 0 and 2 arms over additional independent seeds using the disclosed carved manifest, scene-only initialization, 7.2-million-presentation budget, and final checkpoints. First assert that every held-out camera setting is absent from training and that paired invariant labels match. Evaluate ten trials per official camera task and the nominal four-suite protocol. Report paired task-bootstrap intervals within each seed, training-seed dispersion separately, and spatial/object results for primary extrapolation and compound shifts. A robust replication would retain positive orbital extrapolation differences across seeds without hiding object-task regressions; frequent sign reversals or gains driven by one suite would narrow the aggregate claim. Interpolation must remain a separately reported endpoint. E02E07E08E09E12E13E16E17E24
8.3 Reading coverage
Visual audit: All nine supplied PDF pages were visually inspected after complete reading of all five text chunks. Page 1 verifies title, authors, affiliation and v1 date; p. 2 supplies Figure 1 and release wording; p. 3 supplies the target sequence, action convention and loss; p. 4 supplies proofs, controlled ratios and Figure 2; p. 5 supplies camera ranges, training and evaluation settings; p. 6 supplies Figures 3–4 and Table I; p. 7 supplies Table II, second-seed summaries and task heterogeneity; p. 8 supplies limitations, conclusion and the unresolved reference identifier; p. 9 completes the references. Every final crop was inspected. The architecture's excluded orange link agrees with Eq. (1); the Figure 2 caption qualification and Figure 4 severity-direction ambiguity are disclosed locally. No appendix or complete second-seed table is present; separate supplements and cited works were not inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9. Appendix coverage: not present.
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
- I. Introduction
- II. Related Work
- III. Problem Setup
- IV. Selective Cross-View Consistency, A–C, including inline proofs
- V. A Carve-and-Hold-Out Evaluation Protocol, A–C
- VI. Experiments, A–D
- VII. Discussion
- VIII. Limitations
- IX. Conclusion
- References
Outside the original text pass
- Reviewed revision: arXiv:2608.21402v1 [cs.RO], 7 August 2026. Title and all four authors match the catalog; no other revision was supplied for comparison.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The extraction-only visual gap has been addressed by inspecting all nine PDF pages and all six final crops.
- Separate supplemental material availability has not been fully verified.
- No appendix appears in the supplied nine-page paper. The complete second-seed twelve-bucket table is absent; only main-text summaries are available.
- Code and linked works were not inspected; experiments were not reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
E01PDF p. 1, title, author block, and arXiv footer
The title and authors Bingqi Huang, Bingchuan Wei, Yingkai Cai, and Zhaokui Wang match the catalog. The affiliation is Tsinghua University. The supplied artifact identifies itself as arXiv:2608.21402v1 [cs.RO], dated 7 August 2026.
Go to primary source ↓E02PDF p. 3, Section III, including Assumption 1
Defines same-state camera pairs, invariant actions/proprioception/value, camera-dependent future scenes, the seven-frame sequence, K = 16, and the base-frame action convention with visibility and semantic assumptions.
Go to primary source ↓E03PDF p. 3, Section IV-A, Eq. (1) and following paragraph
Both branches receive supervised denoising loss under shared noise. Consistency selects invariant frames, ramps to 2.0 over 2% of training, applies only to demonstrations, and leaves deployment inputs unchanged.
Go to primary source ↓E04PDF p. 4, Section IV-B, Lemma 1 and Proposition 1 with proofs
The loss separates into mean error and a residual penalty weighted by 1+4λ. Unequal targets yield optimal prediction differences reduced by 1/(1+4λ).
Go to primary source ↓E05PDF p. 4, Section IV-B, Proposition 2 and proof
Independent noise adds twice the cross-covariance trace; at exact invariance this becomes twice the prediction-variance trace. The authors explicitly restrict the positive-sign conclusion.
Go to primary source ↓E06PDF p. 4, Section IV-C and Figure 2 with caption
Synthetic measured ratios are 0.715/0.337/0.116 versus 0.714/0.333/0.111 predictions. At 2,000 steps the 2B wrong-coordinate future-scene guard reaches 0.18 versus approximately 0.98 for selective/control runs; full-scale quantitative verification is deferred.
Go to primary source ↓E07PDF pp. 4–5, Sections V-A and V-B
Training uses 583,648 same-state pairs with explicit azimuth/scale holes and extrapolation bounds; sparse endpoint/elevation training values exclude official held-out settings. Evaluation uses fresh benchmark rollouts rather than training demonstrations.
Go to primary source ↓E08PDF p. 5, Section V-C
Comparison arms share the carved manifest, initialization, and budget and differ in consistency weight. Orbital interpolation/extrapolation are designated primary endpoints with paired task-bootstrap 95% intervals.
Go to primary source ↓E09PDF p. 5, Section VI-A
Describes Cosmos initialization, nominal scene-only reference, 10,000 steps, effective batch 720, 7.2 million presentations, learning rate 5×10^-5, EMA, final checkpoints, ten trials per task, and 500 episodes per nominal suite.
Go to primary source ↓E10PDF p. 5, Section VI-B; PDF p. 6, Figure 3 with caption and Table I
The 1,200-rollout masking conditions score 80.5% unmasked, 66.6% scene-black, 69.7% scene-noise, and 0.2% wrist-black. Table I separately gives released-policy ID/camera success of 98.5/79.5 and scene-only success of 93.2/26.0.
Go to primary source ↓E11PDF p. 6, Section VI-C and Figure 4 with caption
Orbital severity lowers action success from 53% to 3% while relative excess-FID remains between −0.07 and +0.16. The metric subtracts a camera-matched replay floor and measures distributional fidelity rather than semantic plan correctness.
Go to primary source ↓E12PDF p. 7, Table II, C2 azimuth interpolation and extrapolation rows
Interpolation: 49 tasks, control/SCVC 86.3/85.1, difference −1.22 [−4.29,1.84]. Extrapolation: 100 tasks, 63.8/76.0, difference +12.20 [7.40,17.00].
Go to primary source ↓E13PDF p. 7, Section VI-D, second-seed paragraph
Second-seed differences are orbital extrapolation +15.50 [11.70,19.40], dolly +4.44 [0.61,8.49], elevation +4.84 [3.11,6.58], and ID −0.2. Orbital interpolation is −4.29 [−7.56,−1.22], with control/SCVC 90.0/85.7.
Go to primary source ↓E14PDF p. 7, Table II, C1 dolly, C2 elevation, and C3 endpoint rows
Dolly extrapolation: 99 tasks, 67.0/71.2, +4.24 [0.71,7.88]. Elevation: 386 tasks, 54.9/63.7, +8.76 [6.40,11.11]. Endpoint interpolation/extrapolation differences are −0.97 [−3.71,1.61] and −0.58 [−2.12,1.02].
Go to primary source ↓E15PDF p. 7, Table II, ID and C2 azimuth in-distribution rows
Nominal four-suite success is 92.3% control versus 91.7% SCVC. Distribution-matched azimuth improves from 84.0% to 88.5%, with paired difference +4.52 [2.65,6.39].
Go to primary source ↓E16PDF p. 7, Section VII, 'Where the effect stops, and why'
Reports primary-extrapolation spatial gains of 40.8 points, object losses of 11.6, and compound-shift object losses of 40.4. Grasp precision versus viewpoint disorientation is offered as an explanatory interpretation.
Go to primary source ↓E17PDF pp. 7–8, Section VIII
Limits include exact-pair acquisition, simulated WAM evidence, demonstration-only consistency, one elevation extrapolation point, small high-success interpolation buckets, training-seed variation, scene-only scope, and deferred quantitative shrinkage measurement at full scale.
Go to primary source ↓E18PDF p. 2, Introduction, final contribution paragraph
The authors state that code and protocol tooling will be released; this passage does not supply a repository or establish current availability.
Go to primary source ↓E19PDF p. 7, Section VII; PDF p. 8, Section VIII and Reference [7]
Shuffled-pair and real-robot corroboration are attributed to a separate flow-VLA study. Reference [7] contains the unresolved arXiv identifier 2608.XXXXX.
Go to primary source ↓E20PDF p. 2, Figure 1a–b and caption; PDF p. 3, Section III and Eq. (1)
Two same-state views feed a shared denoiser with shared θ and (σ,n). Blue action/proprioception/value outputs are linked by consistency; the orange future-scene link is crossed out and keeps per-view supervision. Panel b shows trained camera coverage, interpolation holes, and extrapolation regions.
Go to primary source ↓E21PDF p. 4, Figure 2 image grid and caption; Section IV-C
Each row compares ground truth, wrong-coordinate predictions, and selective predictions for views A and B. Both wrong-coordinate columns visibly blur scene details; the caption nevertheless describes both arms as tracking the near-nominal view A. The experiment is a short 2,000-step diagnostic, not a converged full-scale quantitative law.
Go to primary source ↓E22PDF p. 6, Figure 4, rotated severity label; Section VI-C, orbital severity paragraph
The rotated 'severity L1 → L5' text points upward on the page, while the prose reports orbital success falling from 53% to 3% across increasing severity with relative excess-FID between −0.07 and +0.16. Individual plotted markers lack explicit level labels.
Go to primary source ↓E23PDF p. 6, Section VI-D; PDF p. 7, Table II; PDF p. 8, Sections VIII–IX
The narrative mentions interpolation counterparts when discussing dolly/elevation extrapolation, but Table II has no elevation interpolation row and Section VIII specifies one elevation extrapolation point. The conclusion says interpolation is intact, whereas the second-seed result has a negative interval excluding zero.
Go to primary source ↓E24PDF p. 5, Section V-C; PDF p. 7, Table II caption and Section VI-D paragraphs immediately below the table
The protocol has twelve axis-level buckets, while Table II shows nine single-axis buckets plus nominal ID. Two compound buckets and the six-task ±2 endpoint bucket are summarized in prose. First-seed compound differences are +0.94 [−2.34,+4.38] and +4.85 [−2.63,+11.82]; the ±2 difference is −5.00 [−13.33,+3.33]. The complete second-seed twelve-bucket table is not present in the supplied PDF.
Go to primary source ↓8.5 Primary sources
Selective Cross-View Consistency for World Action Models: Held-Out Viewpoint Robustness Without Test-Time Camera Information ↗
PDF · 6,994 extracted words
Source fingerprint
5caa5a4c0e9257f15f861563ad188286981b82c47b118929f03e5e712da4d59c