GaussianWAM: Distilling Geometry and Semantics from 3D Gaussian Fields into World-Action Models
1. Paper overview
In one sentence: GaussianWAM uses an offline 3D Gaussian teacher to improve current-observation WAM representations, gaining robustness while preserving the deployed backbone and adding unmeasured teacher-preparation work. e02e03e04e05e06e07e12e16e19
| At a glance | What to know |
|---|---|
| Research problem | Author claim The authors argue that video prediction and language conditioning do not explicitly ensure cross-view geometric consistency or spatially localized object semantics in WAM representations. Independent geometry and semantic losses also lack a shared spatial carrier. GaussianWAM addresses how to organize these signals coherently while retaining the underlying policy's deployment architecture. e02 |
| Core mechanism | |
| A key reported result | FastWAM manipulation under distribution shifts: 71.29 Official overall task success rate (%). FastWAM trained on LIBERO for 70k steps; zero-shot LIBERO-Plus evaluation. FastWAM baseline: 52.05; improvement: 19.24 percentage points. A substantial overall robustness gain in this matched comparison. e12 |
| Reading caution |
Core contributions
Figure 2. One offline teacher construction supports two different WAM architectures. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow panel (a) from agent/wrist images through the geometry and semantic teachers to the Gaussian field, then to cached depth, alpha, feature, and validity maps. The CLIP/DINO/SAM box indicates possible semantic encoders; the experiments use CLIP ViT-B/16. Next compare the lower panels: (b) attaches supervision to current video tokens in a model with separate video and action experts, whereas (c) attaches it to current visual tokens in a shared Transformer. Panel (d) expands the three auxiliary prediction heads. The dashed distillation paths describe training supervision. They do not introduce an additional action decoder at deployment. e20e03e04e05e06e07e08e09e10
What it supports. The same Gaussian targets can shape representations in either architecture without retaining the teacher or its heads at inference. The architecture evidence supports a general training framework: GaussianWAM does not itself determine whether world and action modeling use separate experts or a shared Transformer.
Where the evidence stops. The schematic establishes information flow, not measured speed. Teacher inference, per-sample Gaussian fitting, and caching still require preparation work, whose total time and storage cost are not reported.
2. Motivation
2.1 The problem and the proposed response
The authors argue that video prediction and language conditioning do not explicitly ensure cross-view geometric consistency or spatially localized object semantics in WAM representations. Independent geometry and semantic losses also lack a shared spatial carrier. GaussianWAM addresses how to organize these signals coherently while retaining the underlying policy's deployment architecture. e02
2.2 What this reading follows
A policy can predict plausible video while retaining weak information about object geometry and location. GaussianWAM addresses that gap during training: frozen geometry and semantic teachers supply signals that are bound to common Gaussian primitives, rendered back into aligned targets, and distilled into the policy’s current visual tokens. The field and auxiliary heads disappear at deployment. Read the evidence in two stages: the matched backbone comparisons establish overall benefit, while direct CLIP + VGGT distillation tests how much the Gaussian organization adds. The latter gain is smaller. Physical-task results and depth probes broaden the evidence, but leave uncertainty about general robustness and independent geometric accuracy. e02e03e04e05e06e07e12e16e19
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 | Not applicable |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded outside-quadrant classification fits GaussianWAM as an auxiliary training framework rather than a distinct future/action prediction architecture. Its Gaussian mechanism supervises current visual representations and is removed at inference. Architecture 'Not applicable' is appropriate at this framework level: the FastWAM instantiation has separate interacting experts, whereas Cosmos uses a shared Transformer. The evidence supports 3D multiview supervision and generalization evaluation; action alignment occurs indirectly through backbone representations, without a new alignment decoder or inverse-dynamics mechanism. e03e06e07e08e09e12e13
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 Bind geometry and semantics before asking the policy to learn them
Start with synchronized views of the current scene. VGGT-Omega supplies depth, confidence, and estimated cameras; CLIP ViT-B/16 supplies patch features. The method aligns them on a 14 × 14 teacher grid, scales camera intrinsics, and projects semantic features to 64 dimensions. Every geometry-valid location initializes a Gaussian whose center comes from back-projected depth and whose semantic attribute comes from the corresponding image feature. The field is then fitted jointly across observed views for 50 iterations, using semantic, depth, coverage, and regularization terms. Rendering sends these jointly organized signals back to image-aligned grids. The mechanism therefore depends on a shared set of spatial primitives, not merely the presence of two teachers. Semantic Gaussian features may optionally be refined; the paper does not resolve that implementation choice for the reported runs. e04e05e06
5.2 Follow the training signal into an unchanged action pathway
The rendered field supplies a cache of semantic, depth, alpha, and validity maps. Validity intersects acceptable geometry with rendered coverage above 10⁻⁴, so negligible Gaussian support does not receive auxiliary supervision. Three heads read final-layer tokens for current visual observations: semantic prediction uses cosine distance, while depth and coverage use absolute errors over valid locations. These losses supplement the original WAM objective. In FastWAM, the supervised tokens belong to the video expert and become available to the action expert through MoT interaction. In Cosmos Policy, they belong to the shared diffusion Transformer. After training, teachers, fitting, rendering, and prediction heads are removed. Action generation follows the original policy procedure; this paper does not introduce a Gaussian planner or specify a new feedback schedule. e06e07e08e09
5.3 Separate the large teacher gain from the smaller Gaussian gain
Reader analysis: the headline FastWAM improvement combines several changes and cannot all be attributed to Gaussian organization. LIBERO-Plus success rises from 52.05% to 71.29%, yet direct CLIP + VGGT distillation already achieves 69.37%. Gaussian targets without alpha supervision achieve 70.99%, and the full model adds another 0.30 percentage points. Thus the reported table supports an incremental Gaussian benefit, while its reliability across seeds remains unresolved. Component removal makes depth the largest tested contribution; layer ablation favors final layer 30 over layers 10 and 20, whose overall scores are 67.86% and 68.44%. Those results guide a reproduction of this configuration. They do not establish universal layer placement, immunity to distribution shifts, or ground-truth geometric accuracy: Cosmos regresses on two shifts and the depth probe uses teacher pseudo-labels. e12e13e16e17e18e19
5.4 Training and inference
During training
VGGT and CLIP are frozen. Gaussian fitting precedes policy training; semantic Gaussian features may optionally be refined during fitting. Cached targets avoid repeating teacher inference and field optimization. e04e05e06
FastWAM uses 70k steps on eight A100 GPUs, effective batch size 32, learning rate 10^-4 with cosine decay, and weight decay 10^-2. Its semantic, depth, and alpha weights are 0.01, 0.01, and 0.005. Cosmos uses 5k iterations on eight GPUs with effective batch size 1920. Each enhanced model matches its own baseline's data, configuration, and optimization budget. e10
During inference
Remove foundation teachers, Gaussian construction/rendering, and auxiliary heads. The trained backbone retains its original forward path; the paper does not detail a new action sampling, replanning, or feedback schedule. e03e07e08e09
Removing these modules supports the claim of no additional Gaussian forward computation. The supplied experiments do not provide a measured latency comparison. e03e10
5.5 Implementation flow
- Extract frozen teacher signals
VGGT-Omega estimates depth, confidence, and camera parameters. Depth and confidence are resized to a 14 × 14 grid with adjusted intrinsics. CLIP ViT-B/16 patch features are aligned to that grid and projected to 64 dimensions. e04
- Construct and fit a shared field
Every geometry-valid grid location initializes a Gaussian carrying a 3D center, rotation, scale, opacity, and semantic feature. Depth-aware, opacity-weighted splatting supports joint fitting across observed views for 50 iterations, using semantic, depth, coverage, and regularization terms. e05
- Cache aligned supervision
Render the fitted field back into observed views. Intersect geometry validity with sufficient rendered coverage, then resize and compose targets to match the policy's multi-view token layout. Cache these maps for reuse during policy training. e06
- Distill current visual tokens
Attach semantic, depth, and sigmoid coverage heads to final-layer current-observation tokens. Use masked cosine distance for semantics and masked absolute errors for depth and coverage, adding the resulting objective to the original WAM loss. e07
- Connect supervision to action learning
FastWAM supplies tokens from its video expert, which interacts with the action expert through MoT. Cosmos Policy supplies current visual tokens from its shared Transformer. The auxiliary targets describe current observations rather than future 3D rollouts. e08e09e06
6. Experiments & results
GaussianWAM strengthens existing world-action models by fitting an offline Gaussian field that binds estimated geometry to CLIP features, then teaching current-observation tokens to predict its rendered outputs. The field disappears at deployment. FastWAM's LIBERO-Plus success rises from 52.05% to 71.29%, although direct CLIP and VGGT distillation already reaches 69.37%; the distinctive contribution is therefore the incremental benefit of spatially organizing teacher signals.
6.1 Read the original evidence
Table 1. Both backbones improve overall, while individual shift results remain uneven. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the bottom two blocks, which pair each backbone with its GaussianWAM variant. The LIBERO column is standard evaluation; Camera through Layout are LIBERO-Plus distribution shifts, and Overall is the reported benchmark aggregate. Every entry is a success percentage, while parenthesized changes are percentage-point differences. The caption specifies LIBERO training and zero-shot evaluation on LIBERO-Plus. FastWAM pairs use 70k optimization steps; Cosmos pairs use 5k iterations. Compare within each pair before comparing across model families. Emb. PT. records large-scale embodied pretraining, and dashes in the literature rows denote unreported entries rather than zero success. e10e11e12e13
What it supports. FastWAM improves from 52.05% to 71.29% overall, a 19.24-point gain; Cosmos improves from 71.52% to 77.30%, a 5.78-point gain. Standard LIBERO increases are smaller: 96.60% to 97.60% and 98.50% to 98.60%. These comparisons support benefit on both tested architectures under the reported budgets.
Where the evidence stops. Cosmos background success drops from 77.51% to 66.17%, and lighting from 90.89% to 89.84%. Overall improvement therefore does not imply uniform robustness. The table supplies no uncertainty estimates or universal matched-budget comparison across literature models.
Table 2. RoboTwin transfer improves some scores, but randomized scenes remain difficult. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row as average success percentage on the 15-task RoboTwin 2.0 Clean-to-Random benchmark. The two columns represent distinct evaluation conditions; Random introduces stronger scene and visual randomization. First compare FastWAM with the immediately following enhanced row, then repeat for Cosmos Policy. Keep Clean and Random separate when assessing benefit, since a large improvement in one setting can coexist with low performance in the other. The DP, ACT, and DP3 rows provide additional reference scores. The bold cells identify the best displayed values in each column, but do not indicate statistical significance. e14e10
What it supports. FastWAM preserves 70.50% Clean success and rises from 0.70% to 1.60% under Random. Cosmos rises from 34.40% to 68.90% Clean and from 7.10% to 14.40% Random. This supports positive transfer trends beyond LIBERO, with much stronger absolute performance in clean scenes.
Where the evidence stops. Even the best Random score is 14.40%, and enhanced FastWAM remains at 1.60%. The aggregate table omits per-task outcomes and uncertainty, so it cannot establish broadly reliable bimanual transfer.
Figure 3. Physical execution improves over FastWAM in each displayed condition. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The left images show two tasks performed with two UR7e arms: contacting a red plate with a vase, and placing a cup before inserting a purple cylinder. Notice the differing condition order: the image columns run Standard, Layout, Camera, Noise, while the bar chart runs Standard, Camera, Noise, Layout. Use the legend rather than image position to read gray π0.5, blue FastWAM, and green GaussianWAM bars. The vertical axis is success percentage. The accompanying protocol specifies 100 demonstrations per task and 20 independent evaluation trials per task per setting, with success requiring completion of all task-specific objectives. e15e21
What it supports. GaussianWAM’s Standard/Camera/Noise/Layout bars are 55/30/40/35%, versus FastWAM’s 40/20/30/30%. The prose reports average success increasing from 30% to 40%. These are physical task outcomes, extending the evidence beyond simulated success and visual representation probes. They cover standard scenes and three deliberately shifted evaluation conditions.
Where the evidence stops. The chart has no error bars or separate task-level scores. GaussianWAM ties π0.5 on Camera and trails it on Layout, so its advantage is not universal even within this small physical evaluation.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Standard language-conditioned manipulation LIBERO evaluation after LIBERO training; matched baseline/enhanced training budgets within each backbone. | FastWAM + GaussianWAM: 97.60; Cosmos Policy + GaussianWAM: 98.60. Task success rate (%) | Corresponding baselines: 96.60 and 98.50. Reported gains are 1.00 and 0.10 percentage points under high baseline success. e10e11 |
| FastWAM manipulation under distribution shifts FastWAM trained on LIBERO for 70k steps; zero-shot LIBERO-Plus evaluation. | 71.29 Official overall task success rate (%) | FastWAM baseline: 52.05; improvement: 19.24 percentage points. A substantial overall robustness gain in this matched comparison. e12 |
| Cosmos manipulation under distribution shifts Cosmos Policy trained on LIBERO for 5k iterations; zero-shot LIBERO-Plus evaluation. | 77.30 Official overall task success rate (%) | Baseline: 71.52; improvement: 5.78 percentage points. Overall improvement does not cover every shift: background success falls from 77.51 to 66.17, and lighting from 90.89 to 89.84. e13 |
| FastWAM bimanual manipulation on RoboTwin FastWAM on the 15-task RoboTwin 2.0 Clean-to-Random benchmark. | Clean: 70.50; Random: 1.60. Average task success rate (%) | Baseline Clean: 70.50; Random: 0.70. Clean performance is unchanged; randomized performance improves but remains near failure. e14 |
| Cosmos bimanual manipulation on RoboTwin Cosmos Policy on the 15-task RoboTwin 2.0 Clean-to-Random benchmark. | Clean: 68.90; Random: 14.40. Average task success rate (%) | Baseline Clean: 34.40; Random: 7.10. Both settings improve substantially relative to this baseline, but randomized success remains limited. e14 |
| Physical bimanual manipulation Two UR7e arms; vase-to-plate contact and cup-placement/cylinder-insertion tasks; 100 demonstrations per task and 20 trials per task per setting across standard, layout, camera, and noise conditions. | Reported average: 40.00. Figure 3 Standard/Camera/Noise/Layout: 55/30/40/35. Reported average task success rate (%) | FastWAM average: 30.00; per-setting: 40/20/30/30. π0.5 per-setting: 50/30/35/40. Executed tasks require all objectives to succeed. GaussianWAM improves every setting over FastWAM, but ties π0.5 on Camera and trails it on Layout. No uncertainty is shown. e15e21 |
| Gaussian-field unification ablation FastWAM on LIBERO-Plus; official overall score over 10,030 evaluation tasks. | GaussianWAM without alpha supervision: 70.99; full model: 71.29. Overall task success rate (%) | Direct CLIP + VGGT distillation: 69.37; original FastWAM: 52.05. Most improvement already comes from direct teacher supervision. Field construction without alpha adds a reported 1.62 percentage points; the full model adds 1.92 points over direct distillation, calculated from the reported scores. e16 |
| Supervision-component ablation FastWAM on LIBERO-Plus; official 10,030-task evaluation. | Without semantics: 68.06; without depth: 66.28; without alpha: 70.99. Overall task success rate (%) | Full model: 71.29. Depth removal produces the largest observed drop. Alpha's incremental improvement is small, and no uncertainty is supplied. e17 |
| Supervision-location ablation FastWAM + GaussianWAM on LIBERO-Plus; official 10,030-task evaluation. | Layer 10: 67.86; layer 20: 68.44; final layer 30: 71.29. Overall task success rate (%) | Final-layer supervision exceeds both tested intermediate-layer alternatives. This supports the selected attachment point for this backbone, without establishing a universal layer-selection rule. e18 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. The direct teacher baseline explains most of the headline gain. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the Overall column to compare three stages: original FastWAM, direct CLIP + VGGT supervision, and the full Gaussian model. Then compare the three component-removal rows against Full. The caption defines Overall over the official 10,030 evaluation tasks; it should not be replaced by a simple average of the seven displayed shift percentages. The direct row uses separate semantic and geometry targets without a Gaussian field. The without-alpha row tests Gaussian supervision without that auxiliary component. Inspect individual columns after the aggregate: Full is not the best row for every shift, and boldface marks displayed maxima rather than confidence. e16e17e06e07
What it supports. Direct distillation already reaches 69.37% versus 52.05% for FastWAM. Gaussian targets without alpha reach 70.99%, and Full reaches 71.29%: calculated increments of 1.62 and 1.92 points over direct supervision. Removing depth lowers success to 66.28%, the largest component-removal drop. Adding alpha contributes 0.30 points in this comparison.
Where the evidence stops. No seed variability establishes the stability of these increments. The comparison also does not separately isolate Gaussian fitting, rendering, spatial alignment, and validity handling; matched controls are needed before attributing the entire increment to one mechanism.
Figure 4. Enhanced frozen representations yield depth maps that look closer to the teacher. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The left column supplies the scene reference: RGB above and VGGT Depth below. The middle column contains probes from the baseline FastWAM and Cosmos Policy, and the right column contains the corresponding GaussianWAM versions. Compare each enhanced panel both with its baseline neighbor and with the VGGT panel. The paper freezes each trained WAM backbone and trains the same lightweight probe on final-layer visual representations using VGGT-Omega pseudo-depth supervision. Thus this diagnostic asks what depth information a probe can recover from the representation. It does not show an additional depth module used for deployed action generation. e19e07
What it supports. In this example, the enhanced maps reproduce the teacher’s broad scene structure more closely, especially relative to the predominantly dark Cosmos baseline. The visual supports improved decodability of teacher-like depth from the frozen representation, consistent with the authors’ proposed geometric explanation.
Where the evidence stops. There is no calibrated color scale, numerical depth-error measure, or independent ground-truth panel. Similarity to the same geometry teacher used for supervision cannot by itself demonstrate more accurate physical geometry or explain the action-success gains.
7. Analysis & limitations
7.1 What the evidence leaves open
The unification comparison supports a modest additional gain beyond direct teacher supervision. It does not isolate which aspect of Gaussian fitting or rendering causes that gain, and reported tables provide no seed variability or confidence intervals. e16e17
Robustness remains uneven: Cosmos loses performance under background and lighting shifts, while RoboTwin Random success remains low. Aggregate improvement should not be read as broad distribution-shift reliability. e13e14
Teacher construction depends on synchronized views and estimated geometry. Confidence and coverage filtering restrict unsupported supervision but cannot independently establish that retained geometry is correct. e04e06
The inspected depth example resembles VGGT pseudo-depth more closely after enhancement, but provides no calibrated color scale or quantitative error. It tests teacher agreement, not independent ground-truth geometry. Unlabeled t-SNE groups do not establish semantic accuracy. e19
Deployment architecture is preserved, but offline teacher construction adds work whose total time and cache-storage cost are not reported in the supplied setup. e05e06e10
7.2 Questions for discussion
- Does the incremental Gaussian-field gain persist across seeds when direct distillation uses matched validity handling and head capacity? [e07, e16]
- Does stronger agreement with teacher depth translate into improved geometry against independent ground truth? [e19]
- What explains the Cosmos background regression and the remaining RoboTwin Random failures? [e13, e14]
8. Reproducibility audit
8.1 Requirements and known gaps
Reconstruction requires backbone configurations, synchronized observations, VGGT-Omega and CLIP checkpoints, camera-aware grid alignment, Gaussian fitting/rendering, and cached-target integration. A code link is printed on the title page. e01e04e05e06e10
The text leaves the geometry-confidence cutoff, semantic projection construction, fitting optimizer/weights, exact semantic-refinement choice, auxiliary-head architecture, and outer distillation weight λG unspecified. Coverage threshold 10⁻⁴ is specified and distinct from geometry confidence. Detailed task-level evaluation information remains incomplete. e04e05e06e07e10
A focused replication should compare direct CLIP + VGGT distillation, Gaussian targets without alpha, and the full model under matched FastWAM settings and multiple seeds. Preserve the official LIBERO-Plus aggregation, and measure teacher preparation cost alongside success to test the specific value of field construction. e10e16e17
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: Does spatial unification help after matching the direct-distillation control?
Reader-proposed check, not an executed experiment: train FastWAM with direct CLIP + VGGT targets, Gaussian targets without alpha loss, and full GaussianWAM under the reported 70k-step configuration. Match observations, checkpoint, head capacity, loss scaling, and optimization seeds; add a direct-target control with the Gaussian variant’s validity mask to separate spatial organization from sample selection. Evaluate all variants on the same official 10,030 LIBERO-Plus tasks and report paired seed differences, shift-level scores, cache size, and teacher preparation time. A consistently positive Gaussian-versus-mask-matched-direct gap would support the proposed spatial mechanism. A gap that disappears after matching masks or varies in sign would weaken attribution of the reported 1.62–1.92-point increment. e06e07e10e16e17
Check 2: Does teacher-depth agreement become independently accurate geometry?
Reader-proposed check, not an executed experiment: freeze base and GaussianWAM backbones and fit identical depth probes on disjoint simulator training observations with equal budgets. Compare probes trained on VGGT pseudo-depth with probes trained on simulator ground-truth depth, then evaluate both against held-out ground truth under matched and shifted camera views. Fix the depth scale convention and valid-pixel set, and report absolute depth error alongside object-boundary errors and task success. Improved teacher agreement without lower independent depth error would challenge the geometric-grounding interpretation. Lower ground-truth error across views, especially where the teacher itself is imperfect, would provide stronger support than the single qualitative example. e04e12e19
8.3 Reading coverage
Visual audit: Visually inspected the title, all nine authors, affiliations, and v1/date on p. 1; Figure 1 on p. 2; Figure 2 on p. 4; teacher construction and equations on pp. 5–6; backbone integration and Table 1 on p. 7; training setup and Table 2 on p. 8; Figure 3, real-world protocol, and ablation discussion on p. 9; Tables 3–4, Figures 4–5, and qualitative analysis on p. 10. All six final crops were inspected. Figure 1, Table 4, and Figure 5 were read on full pages but are not standalone crops. All 13 pages were read through the four supplied text chunks; no appendix is present. Separate supplemental material, linked code, and external resources were not inspected.
PDF pages inspected for this edition: 1, 2, 4, 5, 6, 7, 8, 9, 10. 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
- Title, author list, abstract, and version identifier
- 1 Introduction
- 2 Related Work
- 3 Method
- 3.1 Overview
- 3.2 Gaussian Teacher Construction
- 3.3 Gaussian Distillation
- 3.4 FastWAM-Style Dual-Expert MoT
- 3.5 Cosmos-Policy-Style Unified DiT
- 4 Experiments
- 4.1 Experimental Setup
- 4.2 Main Results
- 4.3 Ablation Studies
- 4.4 Qualitative Analysis
- 5 Conclusion
- References
Outside the original text pass
- The supplied artifact is arXiv:2608.24714v1, dated 25 August 2026. Its exact title and all nine authors match the catalog; no revision or edition discrepancy was observed. No other version was supplied or compared.
- 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.
- The extraction-only image limitation was resolved by visually inspecting PDF pages 1, 2, 4, 5, 6, 7, 8, 9, and 10, including all five figures and four tables. All four text chunks, covering all 13 pages, were read.
- No appendix or separate limitations section appears in the supplied PDF. No supplement was supplied.
- Linked code, project pages, and cited publications were not inspected. No experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block, author list, arXiv version line, and resource links
Title, identifier, and all nine listed authors match the catalog. The supplied artifact identifies arXiv:2608.24714v1, dated 25 August 2026, and prints code and project links.
Go to primary source ↓e02PDF pp. 1–3, Section 1 Introduction
The motivation concerns insufficient explicit geometry and localized semantics in video latents, heterogeneous independent teacher signals, and deployment costs of added 3D/4D machinery.
Go to primary source ↓e03PDF p. 4, Section 3.1 Overview
An offline Gaussian teacher supervises WAM visual representations. Teachers, Gaussian modules, and auxiliary heads are removed for inference in both tested architectures.
Go to primary source ↓e04PDF p. 5, Section 3.2, Geometry extraction and Visual-semantic feature extraction, Eq. (1)
Frozen VGGT-Omega provides depth, confidence, intrinsics, and extrinsics. Frozen CLIP ViT-B/16 features are projected to 64 dimensions on a common 14 × 14 grid. Teacher input is visual; language and proprioception remain policy inputs.
Go to primary source ↓e05PDF p. 5, Section 3.2, Gaussian field initialization and Multi-view Gaussian fitting, Eqs. (2)–(4)
Stride-one valid-grid initialization produces geometry/semantic primitives. Multi-view fitting uses depth-aware weighted rendering, semantic cosine and depth absolute errors, coverage and parameter regularization, with 50 fitting iterations and optional semantic refinement.
Go to primary source ↓e06PDF p. 6, Section 3.2, Offline teacher cache, Eqs. (5)–(6)
The cache contains semantic, depth, alpha, and validity maps. Validity intersects geometry validity with coverage above 10^-4. Targets are resized/composed to the policy layout and reused offline.
Go to primary source ↓e07PDF p. 6, Section 3.3 Gaussian Distillation, Eqs. (7)–(13)
Three heads predict semantic features, depth, and sigmoid alpha from final current-observation tokens. Valid-location cosine and absolute-error losses augment the original WAM objective; added modules are removed at inference.
Go to primary source ↓e08PDF p. 7, Section 3.4 FastWAM-Style Dual-Expert MoT, Eq. (14)
Current tokens are selected from final video-expert hidden states. MoT interaction makes the supervised video representation available to a separate action expert without adding an action pathway.
Go to primary source ↓e09PDF p. 7, Section 3.5 Cosmos-Policy-Style Unified DiT, Eq. (15)
Cosmos processes visual, world-related, and action-related tokens in a unified diffusion Transformer. Gaussian losses target its final current-observation visual tokens.
Go to primary source ↓e10PDF pp. 7–8, Section 4.1 Experimental Setup, especially Backbones and training
The setup supplies FastWAM's 70k-step optimizer and batch configuration and loss weights, Cosmos's 5k iterations and effective batch 1920, and matched data/configuration/budgets for each baseline pair.
Go to primary source ↓e11PDF p. 7, Table 1, FastWAM and Cosmos Policy baseline/enhanced rows, LIBERO column
LIBERO success is 96.60 to 97.60 for FastWAM and 98.50 to 98.60 for Cosmos Policy.
Go to primary source ↓e12PDF p. 7, Table 1 caption and FastWAM baseline/enhanced rows, Overall column; p. 8, Section 4.2
Models train on LIBERO and undergo zero-shot shifted evaluation on LIBERO-Plus. FastWAM overall success rises from 52.05 to 71.29, a reported 19.24-point gain.
Go to primary source ↓e13PDF p. 7, Table 1, Cosmos Policy baseline/enhanced rows, Overall, Light, and BG columns
Overall success rises from 71.52 to 77.30. Lighting changes from 90.89 to 89.84 and background from 77.51 to 66.17.
Go to primary source ↓e14PDF p. 8, Table 2 caption and FastWAM/Cosmos rows; Section 4.2 RoboTwin discussion
The 15-task RoboTwin 2.0 table gives FastWAM Clean/Random as 70.50/0.70 versus 70.50/1.60, and Cosmos as 34.40/7.10 versus 68.90/14.40.
Go to primary source ↓e15PDF p. 9, Section 4.2, Real-world manipulation paragraphs
Two UR7e arms perform two specified tasks with 100 demonstrations each and 20 independent trials per task per setting. Success requires all objectives. The prose reports average FastWAM success of 30.00 versus GaussianWAM's 40.00.
Go to primary source ↓e16PDF p. 9, Section 4.3 Effect of Gaussian-field unification; p. 10, Table 3 caption and Overall column
Direct CLIP + VGGT gives 69.37, GaussianWAM without alpha 70.99, and full GaussianWAM 71.29, against FastWAM's 52.05. The text reports a 1.62-point unification gain without alpha; overall evaluation covers 10,030 tasks.
Go to primary source ↓e17PDF p. 10, Table 3, Full and component-removal rows, Overall column; p. 9, component-ablation discussion
Full success is 71.29; removing semantics gives 68.06, depth 66.28, and alpha 70.99. No variability estimates accompany these scores.
Go to primary source ↓e18PDF p. 10, Table 4 caption and Overall column; pp. 9–10, Effect of supervision depth
Layers 10, 20, and 30 achieve 67.86, 68.44, and 71.29 overall on 10,030 tasks. Layer 30 is the final Transformer layer.
Go to primary source ↓e19PDF p. 10, Figures 4–5 and Section 4.4 Qualitative Analysis
Figure 4 shows RGB, VGGT Depth, and base/enhanced FastWAM and Cosmos depth probes. Frozen backbones use the same lightweight probe with VGGT-Omega pseudo-depth supervision; enhanced examples resemble the teacher more closely. No calibrated depth scale or quantitative error is shown. Figure 5 shows four t-SNE panels without a semantic-group legend; the authors describe more compact and separated groups using the same projection configuration.
Go to primary source ↓e20PDF p. 4, Figure 2, panels (a)–(d) and caption; p. 5, Section 3.2
Panel (a) constructs/caches a training-only Gaussian teacher; (b) attaches supervision to current video representations in a dual-expert model; (c) uses shared Transformer visual tokens; (d) shows semantic, depth, and alpha heads. DINO and SAM appear as schematic alternatives, while Section 3.2 states CLIP ViT-B/16 is used in all experiments.
Go to primary source ↓e21PDF p. 9, Figure 3, right-panel legend and Standard/Camera/Noise/Layout bars; caption and real-world evaluation paragraphs
Gray π0.5 bars read 50/30/35/40%, blue FastWAM 40/20/30/30%, and green GaussianWAM 55/30/40/35%. Scene columns on the left instead run Standard/Layout/Camera/Noise. Each task has 100 demonstrations and 20 trials per setting. No error bars or separate task-specific scores appear.
Go to primary source ↓8.5 Primary sources
GaussianWAM: Distilling Geometry and Semantics from 3D Gaussian Fields into World-Action Models ↗
PDF · 6,434 extracted words
Source fingerprint
0bd4e1cbbeba497be058a3411fbee550c847a629f96f701e394f4e19568cebd5