4DGS-WAM: Bridging Past and Future with an Object-Centric World Action Model based on 4D Gaussian Splatting
1. Paper overview
In one sentence: 4DGS-WAM forecasts actor motion and transports existing object Gaussians, gaining persistent scene structure while inheriting perception, camera and unseen-content limitations. e-overviewe-perceptione-worlde-futuree-reconstructione-limits
| At a glance | What to know |
|---|---|
| Research problem | Source description The paper targets redundant background generation and implicit object geometry in image-based world action models. It asks whether accumulated 4D structure can support future rendering by updating dynamic objects while reusing observed static content. e-overview |
| Core mechanism | Source description An object-centric interface joins reconstructed Gaussian geometry, object associations and center trajectories to separately trained policy and world networks. e-overviewe-perceptione-training |
| A key reported result | Future RGB prediction at the given camera: 18.80±1.21; 0.596±0.091; 0.161±0.042 PSNR ↑ (dB); SSIM ↑; LPIPS ↓. KITTI-MOT; identical prefixes and targets; h=1,3, six frames per method. Mean ± standard deviation over three sequence-level means; sequence IDs and train/test split are unspecified. Epona: 17.46±1.21; 0.482±0.054; 0.203±0.021. KITTI-adapted DriveDreamer-2: 13.73±2.67; 0.360±0.082; 0.426±0.064. Best reported given-camera row on all metrics. The frozen underlay is part of 4DGS-WAM's output, while baselines use native outputs; this system comparison does not isolate dynamics gains. e-futuree-evaluation-scope |
| Reading caution | Source description Per-object dynamics cannot model collisions or instantiate newly appearing objects. The state cannot complete unseen object sides or newly revealed background; perception errors propagate. Deformable manipulation, long-horizon prediction and unified end-to-end modeling remain future work. e-limits |
Core contributions
- Source description
An object-centric interface joins reconstructed Gaussian geometry, object associations and center trajectories to separately trained policy and world networks. e-overviewe-perceptione-training
Figure 1. A persistent scene connects observed images to predicted object motion. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the arrows from left to right. Past RGB is processed by 4DGS reconstruction, segmentation, depth/pose estimation and optical flow. These complementary estimates form the state s_t: observed Gaussians, object trajectories and object associations. The policy π uses the trajectory information to predict actor motion; the world model W receives the observed state and the predicted action. The second blue box represents the future scene, containing both transformed object splats and background splats. Rendering produces the final RGB view. Section 3.1 clarifies that current object masks associate splats with dynamic objects, while the remaining Gaussians form the persistent static bank. e-overviewe-perceptione-worlde-futuree-limits
What it supports. The key architectural choice is to predict changes to an explicit scene. Previously observed background content remains available while the learned transition focuses on object splats. This provides a concrete representation of scene persistence, but the diagram alone establishes neither an efficiency gain nor an improvement in physical reasoning.
Where the evidence stops. The static bank retains observed content; it cannot generate newly revealed background. Section 6 also excludes newly appearing objects and shape completion. The quantitative render pipeline adds a frozen fusion underlay to fill remaining holes.
2. Motivation
2.1 The problem and the proposed response
The paper targets redundant background generation and implicit object geometry in image-based world action models. It asks whether accumulated 4D structure can support future rendering by updating dynamic objects while reusing observed static content. e-overview
2.2 What this reading follows
A driving video contains moving cars, a moving camera and much background that persists across frames. 4DGS-WAM turns that distinction into its representation: observed images become an object-associated Gaussian scene, a policy predicts actor motion, and a separate world model updates dynamic splats while retaining static content. This illustrated reading follows the state from reconstruction to future rendering, then examines what the KITTI-MOT comparisons actually test. Camera choice matters, and the reported images include a frozen fusion underlay. The strongest reconstruction scores concern already observed views; the short-horizon prediction evidence does not establish collision handling or closed-loop driving. e-overviewe-perceptione-worlde-futuree-reconstructione-limits
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 | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is WAMs. View the current classification.
3.1 Evidence-based assessment
Conflicts with the recorded classification
The foundational/world-simulator and 3D-state categories fit the topic. However, architecture and prediction-paradigm 'Not applicable' labels underdescribe the modular policy→world-model design and action-conditioned forward Gaussian transport. Figure 2 distinguishes it from unified image/action prediction; separate training confirms distinct networks. This is neither a demonstrated One Model architecture nor inverse dynamics. A specific quadrant is not assigned without its catalog definition. e-overviewe-worlde-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 Build a scene memory before predicting motion
Begin with what the state remembers. The reconstructed 4DGS contains spatial Gaussian attributes together with lifespans and motion twists describing the observed interval. Reconstruction alone does not identify which primitives belong to which moving object. Segmentation, depth, camera pose and optical flow supply the associations and world-space trajectories that make the state object-centric. At the last observed time, masks partition active Gaussians into dynamic-object subsets and a static remainder. The policy consumes center histories; flow correspondences supervise the world model. This distinction also fixes the temporal boundary: target-time masks, centers and correspondences may be extracted from future observations during training, but none is an inference input. Retaining static content preserves observations; it does not complete background that has never been seen. e-perceptione-worlde-limits
Figure 3. Actor prediction and Gaussian transport use different information and separate networks. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each panel upward from its inputs. In panel (a), trajectories, a learned memory token m and horizon h enter temporal attention followed by actor kNN attention. The action head reads the memory feature. Appendix A.2 specifies eight center-state observations and k=4 neighboring actors. In panel (b), observed splats and category-specific part tokens enter geometric layers conditioned by the action and horizon. The geometry branch composes residual motion with the supplied actor transform; the appearance branch mixes residual updates into source attributes. Appendix A.3 makes this concrete: part-node twists are blended onto splats, while local source-color transport contributes to appearance changes. e-policye-worlde-appearancee-evaluation-scope
What it supports. Interactions among nearby actors occur in the policy, whereas W updates each object's Gaussian subset separately. These are distinct modeling choices: trajectory context can influence predicted actor motion, but per-object splat transport does not implement collision response. The two panels therefore should not be read as a unified image/action backbone.
Where the evidence stops. W receives teacher actions during training and policy predictions at inference. The source reports no controlled ablation of that conditioning difference, the part-token counts or residual motion heads, so their individual benefits remain unmeasured.
5.2 Ask for one horizon, then compose object and part motion
A query supplies h, the interval from current time t to target time t′. The policy processes the observed history once, even when h exceeds one. Its action head predicts a bounded correction to a constant-velocity prior in per-step-equivalent coordinates, and the horizon map converts that result into cumulative translation and rotation. Ego motion provides the future camera; other actor actions condition their object updates. The world model then predicts residual twists at part nodes, blends them onto splats and composes them after the supplied action. A car uses one part token as a rigid fallback, while a person uses five. These part counts are architectural prescriptions, not experimentally isolated improvements. At initialization the residual motion decoder is zero, so transport starts from the supplied rigid action. e-policye-actione-worlde-evaluation-scope
5.3 Train transport and appearance without conflating their supervision
The source trains the policy and world network separately. Appendix A.4 is essential because it replaces schematic losses with the implemented normalization, confidence weighting and auxiliary terms. Policy regression compares bounded predictions with teacher residuals normalized around constant velocity. For W, dense-flow supervision asks where matched surface points move; center transport asks whether their mean displacement agrees with teacher translation. Masked RGB error supervises appearance, while missing-alpha loss penalizes unexplained object pixels and a geometry anchor discourages parameter drift. Reader analysis: these constraints act on different failure modes, so a good image score alone cannot identify which one contributed. Another unresolved issue is the conditioning change from teacher actions during training to predicted actions at inference. A controlled comparison is needed before attributing the final rendering gains to learned dynamics. e-traininge-policy-losse-world-losse-worlde-evaluation-scope
5.4 Training and inference
During training
Train separately: policy AdamW, learning rate 3×10⁻⁴, batch 16, 40 epochs, 1.5-epoch warmup and cosine decay to one-tenth peak; world-model Adam, 3×10⁻⁵ for geometry and appearance, batch 4, 10 epochs with cosine decay. Both clip gradients at 5. The world model receives teacher actions, creating a training/inference conditioning difference. e-traininge-world
Appendix A.4 refines the schematic policy loss: masked joint Charbonnier regression operates in normalized residual coordinates, with confidence-weighted actor averaging and ε=10⁻⁶. World geometry uses confidence-weighted coordinatewise Charbonnier flow and center transport; photometry uses target object masks. The appendix supplies the full multi-horizon objective below. e-policy-losse-world-loss
During inference
For each queried h, evaluate the policy once on the observed prefix, then condition the world model on predicted object actions and the same horizon. No intermediate autoregressive actions or target-time perception are required. Render at the given evaluation camera or predicted ego camera. Quantitative outputs fill uncovered pixels with a frozen NeoVerse fusion underlay. No executed control loop or physical action success is evaluated. e-perceptione-policye-futuree-evaluation-scope
5.5 Implementation flow
- Lift and associate observations
A NeoVerse-referenced reconstructor produces Gaussians with position, rotation, scale, opacity, spherical-harmonic appearance, lifespan and reconstructed forward/backward twists. Segmentation, depth, pose and optical flow establish object identities, centers and world-space flow correspondences. Current masks associate active splats with objects; remaining splats form the static bank. Target-time observations supply training supervision only. e-perception
- Predict cumulative actor motion
The policy uses eight 9-D tokens per actor: center, displacement and optional uncertainty, zero when unavailable. Width-96 embeddings with horizon and memory tokens pass through three four-head causal temporal blocks and actor kNN attention with k=4. The memory feature predicts a bounded residual over constant velocity. Ego motion determines the predicted camera. e-policye-action
- Transport parts and appearance
The world model centers each object's splats and samples category-specific part seeds: one for cars and other rigid fallback categories, three for bicycles/motorcycles, five for people. Four width-128 geometric tensor layers combine splat kNN edges (k=16), splat–part and part–part edges. Part twists blend onto splats and compose with the actor action. Source-color transport and bounded color, opacity and scale residuals update appearance. Zero-initialized final heads initially preserve the supplied rigid motion and source appearance. e-worlde-appearance
6. Experiments & results
4DGS-WAM reconstructs a persistent Gaussian scene, predicts actor motions, and transports existing object splats while retaining the static background. KITTI-MOT experiments support short-horizon image prediction and observed-view reconstruction; camera prediction and a frozen fusion underlay materially affect those results.
The complete supplied paper has architecture figures, two quantitative comparisons and qualitative examples, but no controlled ablation of policy residuals, part tokens, loss terms or the fusion underlay. Five visuals cover the available mechanism and results evidence; no ablation visual is invented. Table 2's full-frame fusion and dynamic-region no-fusion columns change the scored region and therefore do not supply the missing matched underlay experiment. No physical-control or long-horizon result is reported. e-evaluation-scopee-reconstructione-limits
6.1 Read the original evidence
Table 1. Camera choice changes the comparison, and the predicted-camera metrics disagree on the winner. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the Camera column before comparing scores. Epona and DriveDreamer-2 use the given future camera; Envision4D uses its predicted camera. The two 4DGS-WAM rows render the same policy-fed scene prediction at different cameras. PSNR and SSIM increase with quality; LPIPS decreases. The retained table note defines the small evaluation: horizons h=1,3, six target frames per method, and mean±standard deviation over three sequence-level means. Bold marks apply only to given-camera rows. The dagger identifies KITTI-adapted DriveDreamer-2 conditioning, so that row is not an evaluation using its native nuScenes inputs. e-futuree-predicted-camerae-evaluation-scope
What it supports. At the given camera, WAM reports 18.80 PSNR, 0.596 SSIM and 0.161 LPIPS, ahead of both video baselines. At predicted cameras, WAM has lower PSNR and SSIM than Envision4D, but better LPIPS: 0.245 versus 0.307. Thus the text's unqualified statement that it trails Envision4D is too broad for the displayed numbers.
Where the evidence stops. Only WAM composites its prediction over the frozen NeoVerse fusion underlay; baselines use native outputs. The table does not isolate the underlay's contribution. Sequence-level variation is not a confidence interval or evidence from repeated training seeds.
Table 2. Observed-view reconstruction is strong, with different fusion settings for full-frame and dynamic-region scores. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The row groups identify three reconstruction families: point clouds, static Gaussian methods and dynamic Gaussian methods. Compare the first three columns for full-frame image quality, then read the two dyn columns separately for dynamic-object regions. The retained note is essential: WAM's full-frame values include a frozen NeoVerse fusion underlay, whereas its dynamic-region values come from no-fusion evaluation. Point-cloud rows contain dashes in the dynamic columns, indicating absent values rather than zeros. Section 4.3 restricts this experiment to observed input views at their given past cameras. These are reconstruction scores, separate from the future-prediction task in Table 1. e-reconstructione-evaluation-scope
What it supports. WAM reports full-frame PSNR 27.63±2.07 against MonoGS's 20.42±5.55. Its dynamic-region PSNR is 25.53±3.58 and dynamic LPIPS 0.025±0.006, with no fusion. All five WAM values lead their reported columns, supporting strong observed-view reconstruction under the stated settings rather than establishing unseen-future accuracy.
Where the evidence stops. Full-frame fusion scores and dynamic-region no-fusion scores use different image regions. Their coexistence is not a controlled underlay ablation. The paper gives no matched full-frame fusion-off row, and reconstruction of observed views cannot establish novel-view generalization.
Figure 4. Qualitative prediction panels mix given and predicted camera protocols. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row as one target scene and each column as a method. The columns are ground truth, Epona, DriveDreamer-2, Envision4D and Ours. The caption assigns horizons h=1, h=2 and h=1 from top to bottom; these examples therefore should not be mistaken for the exact h=1,3 quantitative sample in Table 1. Epona and DriveDreamer-2 use given future cameras, while Envision4D and WAM use predicted cameras. Compare road markings and vehicle positions as well as local sharpness. WAM's static content is reconstructed from the prefix, and its displayed output includes the frozen fusion underlay. e-qualitativee-futuree-limitse-evaluation-scope
What it supports. The examples make the representational idea visible: a predicted image can retain recognizable road and building structure while placing dynamic actors in a future scene. As reader analysis, sharp background detail here is consistent with scene reuse, but these composite images cannot attribute that detail separately to the static bank, object transport or fusion.
Where the evidence stops. Camera differences can produce apparent misalignment independently of object dynamics. These selected panels, resized to native aspect ratio for display where needed, do not establish a quantitative ranking, long-horizon stability or successful physical action execution.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Future RGB prediction at the given camera KITTI-MOT; identical prefixes and targets; h=1,3, six frames per method. Mean ± standard deviation over three sequence-level means; sequence IDs and train/test split are unspecified. | 18.80±1.21; 0.596±0.091; 0.161±0.042 PSNR ↑ (dB); SSIM ↑; LPIPS ↓ | Epona: 17.46±1.21; 0.482±0.054; 0.203±0.021. KITTI-adapted DriveDreamer-2: 13.73±2.67; 0.360±0.082; 0.426±0.064. Best reported given-camera row on all metrics. The frozen underlay is part of 4DGS-WAM's output, while baselines use native outputs; this system comparison does not isolate dynamics gains. e-futuree-evaluation-scope |
| Future RGB prediction at the predicted ego camera Same KITTI-MOT prefixes, targets and aggregation as Table 1; camera predicted by the policy. | 16.47±3.05; 0.482±0.033; 0.245±0.107 PSNR ↑ (dB); SSIM ↑; LPIPS ↓ | Envision4D predicted camera: 17.19±0.65; 0.514±0.072; 0.307±0.069. Lower PSNR and SSIM but better LPIPS than Envision4D. Section 4.2's blanket statement that it trails Envision4D overlooks the LPIPS ordering. Camera prediction degrades all three metrics relative to this model's given-camera rendering. e-predicted-camera |
| Past reconstruction of observed input views KITTI-MOT, given past cameras, mean ± standard deviation over three sequences. Full-frame WAM scores include fusion; its dynamic-region scores use no fusion. | 27.63±2.07, 0.888±0.030, 0.053±0.012; 25.53±3.58, 0.025±0.006 Full-frame PSNR ↑ (dB), SSIM ↑, LPIPS ↓; dynamic-region PSNR ↑ (dB), LPIPS ↓ | MonoGS: 20.42±5.55, 0.652±0.225, 0.236±0.177; 15.51±5.12, 0.224±0.159. Best reported values in Table 2, measuring already observed views. Dynamic-region values provide a separate no-fusion result, not a matched full-frame underlay ablation. e-reconstruction |
6.3 Ablations and diagnostic examples
No ablation or diagnostic visual is included in this edition. The source-specific visual limitations are recorded above.
7. Analysis & limitations
7.1 What the evidence leaves open
Per-object dynamics cannot model collisions or instantiate newly appearing objects. The state cannot complete unseen object sides or newly revealed background; perception errors propagate. Deformable manipulation, long-horizon prediction and unified end-to-end modeling remain future work. e-limits
Three-sequence image metrics and selected panels do not establish generalization, physical interaction accuracy or closed-loop driving. No controlled ablation isolates part tokens, learned policy residuals or fusion. Standard deviations describe sequence variation, not repeated training seeds or confidence intervals. e-futuree-reconstructione-evaluation-scope
7.2 Questions for discussion
- Does the learned policy residual improve over constant velocity after controlling object transport and camera choice?
- How much full-frame quality comes from transported splats, the static bank and the frozen underlay separately?
- What interaction representation could address collisions while preserving persistent object geometry?
8. Reproducibility audit
8.1 Requirements and known gaps
Reconstruction, segmentation, depth/pose, flow, object associations and teacher-action extraction are prerequisites. Preserve the unusual rotation convention R(a)=R_z(roll)R_x(pitch)R_y(yaw), with yaw about the vertical axis; substituting conventional Euler semantics would alter supervision and rendering. e-perceptione-action
Sequence identities, training-set size/split, horizon sampling, hardware, runtime and the precise fusion implementation are unspecified. Proposed checks should hold prefixes and cameras fixed, compare constant velocity with learned policy residuals, and compare rendering with/without fusion on identical masks. Neither check has been run. e-actione-futuree-reconstructione-evaluation-scope
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 the learned policy improve on its constant-velocity prior?
Reader-proposed, not executed: use identical held-out prefixes, horizons and perceived splats, and fix the render camera to the given pose. Compare constant-velocity actions, learned policy actions and teacher actions while holding the trained world model and underlay fixed. Score actor translation/heading error against the same teachers, plus dynamic-region image metrics on identical masks; report each horizon and sequence separately. The learned policy should improve over constant velocity and approach the teacher-action condition. If it does not, the present image evidence cannot attribute gains to learned actor forecasting. A further frozen zero-residual W control can distinguish supplied rigid motion from learned transport, without retraining either network. e-actione-policye-worlde-futuree-evaluation-scope
Check 2: Separate camera error from fusion coverage
Reader-proposed, not executed: render the same predicted Gaussian state in a two-by-two comparison of given/predicted camera and fusion on/off. Hold the state, target images, rendering settings and evaluation masks fixed; explicitly document how uncovered pixels are scored. Report full-frame and dynamic-region PSNR/SSIM/LPIPS, and stratify pixels by splat coverage before adding fusion. The camera effect is the paired change at fixed fusion; the underlay effect is the paired change at fixed camera. If the advantage concentrates in previously uncovered background pixels, it supports a coverage explanation rather than improved dynamic transport. The published dynamic-region no-fusion scores alone cannot answer this matched full-frame question. e-futuree-predicted-camerae-reconstructione-limits
8.3 Reading coverage
Visual audit: Visually inspected the title/author/version page, Figures 1–5, Tables 1–3 and method/training equations on the declared pages. All five final original crops were viewed after extraction. Table 1 retains its combined evaluation-protocol and footnote block because the bold marks and dagger depend on it; Table 2 retains its fusion/no-fusion note. Figure captions and body prose are excluded from figure crops. The experimental pages contain no controlled ablation; Appendix A supplies implementation details only. References on pages 9–10 were read in text and were not visually inspected. No separate supplement was supplied.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 11, 12, 13, 14. 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
- PDF pp. 1–2: title, authors, Abstract and Section 1 Introduction
- PDF pp. 2–3: Section 2 Related Work
- PDF pp. 3–6: Section 3, Past Perception and Future Prediction
- PDF pp. 6–8: Sections 4.1–4.4, implementation, prediction, reconstruction and qualitative evaluation
- PDF p. 8: Sections 5–6, Conclusion and Limitations and Future Work
- PDF pp. 9–10: References
- PDF pp. 11–14: Appendix A.1–A.4, notation, policy, world model and exact training objectives
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.
- The extraction's visual limitation was addressed by inspecting PDF pages 1–8 and 11–14, all five figures, all three tables and appendix equations. References on pages 9–10 were read as text only.
- No separate supplement was supplied. Code was not inspected and experiments were not reproduced.
- Identity/version: title and authors match the catalog after normalizing typography and name order. The title page identifies arXiv:2608.25956v1, 26 August 2026, and labels the preprint a work in progress. No title, author or affiliation discrepancy was found. No other edition was supplied or compared.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title/author block, arXiv stamp and footnote
Title matches observedTitle. Authors are Yueen Ma, Zenglin Xu and Irwin King. Ma: CUHK and Shanghai Academy of AI for Science; Xu: Fudan and the Academy; King: CUHK. Stamp identifies v1, 26 August 2026; footnote says work in progress.
Go to primary source ↓e-overviewPDF pp. 1–3, Abstract, Section 1 and Figures 1–2
Motivates persistent spatial structure and static/dynamic decomposition. Figures show 4D perception feeding separate policy and world networks, unlike the illustrated unified token-based WAM.
Go to primary source ↓e-perceptionPDF p. 4, Section 3.1, Equations 1–2 and final paragraphs
Defines Gaussian attributes, lifespan and twists, object-center histories and 3D flow. Vision estimates give world-frame associations. Future masks/centers/correspondences are training-only; current masks split objects from static content.
Go to primary source ↓e-policyPDF p. 5, Section 3.2 Policy, Equation 3; p. 12, Appendix A.2 Input tokens and attention, Equation 11
One evaluation predicts target-horizon actions including ego motion. Eight 9-D center-state observations feed width-96 embeddings, memory/horizon tokens, three four-head temporal blocks and k=4 actor attention.
Go to primary source ↓e-actionPDF pp. 11–12, Appendix A.2 Action convention, Equations 7–10
Specifies six-vector ordering, R_z(roll)R_x(pitch)R_y(yaw), bounded constant-velocity residual scales and horizon mapping. Teachers use centroid displacement, tangent heading change, zero object roll/pitch and relative ego camera motion.
Go to primary source ↓e-worldPDF p. 4 Figure 3(b); p. 5 Section 3.2 Equation 5; pp. 12–13 Appendix A.3 Equations 12–22
Per-object horizon/action-conditioned geometric network transports dynamic splats and retains static content. Specifies category part counts, graph edges, four width-128 layers, blended twists, composition order and zero initialization. Teacher actions train W; policy predictions condition inference.
Go to primary source ↓e-appearancePDF p. 13, Appendix A.3 Appearance transport, Equations 23–26
Source-color transport and bounded color, opacity-logit and log-scale residuals update appearance. Zero final-layer initialization gives an identity appearance update.
Go to primary source ↓e-trainingPDF p. 6, Section 4.1 Implementation Details
Separate policy/world training uses the stated AdamW/Adam rates, batches, epochs, schedules and clipping. Loss weights add missing alpha and geometry anchoring.
Go to primary source ↓e-policy-lossPDF p. 13, Appendix A.4 Policy loss, Equations 27–28
Exact policy regression uses normalized bounded residuals, masked joint Charbonnier, confidence-weighted averaging and epsilon 10^-6, refining schematic Equation 4.
Go to primary source ↓e-world-lossPDF p. 14, Appendix A.4, Equations 29–33
Confidence-weighted dense/center flow losses use coordinatewise Charbonnier. Object-mask RGB error, missing alpha and geometry anchoring form Equation 33's weighted multi-horizon objective.
Go to primary source ↓e-futurePDF p. 6, Table 1 caption, given-camera rows and Section 4.2
Same prefixes/targets, h=1,3, six frames per method, mean±std over three sequence means. WAM's given-camera PSNR/SSIM/LPIPS are 18.80±1.21/0.596±0.091/0.161±0.042, beating both video rows. DriveDreamer-2 uses KITTI-adapted conditioning. WAM includes frozen NeoVerse fusion; baselines use native outputs.
Go to primary source ↓e-predicted-cameraPDF p. 6, Table 1 predicted-camera rows and Section 4.2 final sentence
WAM PSNR/SSIM/LPIPS: 16.47±3.05/0.482±0.033/0.245±0.107; Envision4D: 17.19±0.65/0.514±0.072/0.307±0.069. Lower-is-better LPIPS favors WAM despite the prose's unqualified 'trails' statement.
Go to primary source ↓e-reconstructionPDF p. 7, Table 2 caption, WAM/MonoGS rows and Section 4.3
Input views only, given past cameras, three-sequence mean±std. WAM full-frame PSNR/SSIM/LPIPS: 27.63±2.07/0.888±0.030/0.053±0.012; dynamic PSNR/LPIPS: 25.53±3.58/0.025±0.006. MonoGS: 20.42±5.55/0.652±0.225/0.236±0.177 and 15.51±5.12/0.224±0.159. WAM uses fusion for full-frame, no fusion for dynamic scores.
Go to primary source ↓e-qualitativePDF p. 7, Section 4.4; p. 8, Figures 4–5 and captions
Figure 4 rows use h=1,2,1; columns GT/Epona/DriveDreamer-2/Envision4D/Ours. Video baselines use given cameras; Envision4D and WAM predicted cameras. Figure 5 shows observed-view reconstruction. WAM panels include fusion; canvas resizing is display-only.
Go to primary source ↓e-limitsPDF p. 8, Section 6 Limitations and Future Work
Lists no collisions, new objects, unseen-side completion or newly revealed background generation; perception dependence, fusion asymmetry, short horizons and untested manipulation/deformability. Unified end-to-end modeling remains future work.
Go to primary source ↓e-evaluation-scopePDF pp. 6–8, Sections 4–6 and Tables 1–2; pp. 11–14, Appendix A.1–A.4
Evaluation comprises future images, observed-view reconstruction and qualitative examples. No controlled mechanism ablation or physical/closed-loop evaluation is reported. Implementation details do not specify sequence IDs, train/test split, training-set count, horizon sampling, hardware/runtime or a full fusion recipe.
Go to primary source ↓8.5 Primary sources
4DGS-WAM: Bridging Past and Future with an Object-Centric World Action Model based on 4D Gaussian Splatting ↗
PDF · 6,715 extracted words
Source fingerprint
5f26d51df62eae84b8fe520a151ce073fc043bce47e03ed4e50189820827d14f