GeoWAM: Visual Geometry World Action Models for Autonomous Driving
1. Paper overview
In one sentence: GeoWAM turns forecast geometry into a conditioning signal for a deterministic driving policy, with strong reported benchmark scores but no controlled test isolating the source of its gains. e01e04e07e08e11e12e14e16
| At a glance | What to know |
|---|---|
| Research problem | Author claim The authors argue that future-image prediction entangles appearance with spatial dynamics, whereas metric point maps expose structure in a space aligned with motion planning. The practical question is whether predicting that geometry provides a better policy representation, beyond producing plausible future observations. e02 |
| Core mechanism | |
| A key reported result | Two-stage planning on NAVSIM v2 navhard: 36.6. Overall EPDMS (higher better). Original scenes in S1, synthetic reactive scenes in S2; 3D Gaussian Splatting pseudo-closed-loop feedback; official human-penalty protocol. DVGT-2: 31.7; DriveLaW: 30.6; EponaV2†: 36.1; NavFormer†: 34.1. Dagger denotes reinforcement-learning or PDMS-score supervision. Highest reported aggregate, including differently supervised baselines. The merged EPDMS cell summarizes the method, not S2 alone. GeoWAM's NC declines from 97.7 in S1 to 80.4 in S2; robustness remains limited. e10e13e14 |
| Reading caution | Reader analysis The supplied experiments contain baseline comparisons and three selected qualitative cases, but no ablation of pretraining, loss terms, stop-gradient or future-token conditioning. Causal attribution to any one mechanism is therefore unresolved; there is no dedicated limitations section. e16 |
Core contributions
Figure 1. The motivating choice: predict spatial structure as the future state. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the multiview observation at the left and follow the two arrows. The upper branch produces a sequence of future images; the lower branch produces a geometric scene with successive ego poses. The right-hand boxes express the authors' representation argument: transformations must be recovered indirectly from changing images, while changes between point sets expose spatial evolution. The image and point labels distinguish prediction spaces, rather than two measured performance curves. Read this schematic alongside the introduction: GeoWAM still starts from RGB observations, but its forecast target is geometric structure instead of future image appearance. e02e05
What it supports. The paper's proposed advantage is a closer alignment between the predicted state and driving motion. This visual explains why geometry is used as the intermediate representation between observation and planning. It motivates the design; it does not itself demonstrate that geometry forecasts are accurate or that their trajectories are safer.
Where the evidence stops. The explicit-versus-implicit distinction is the authors' modeling argument. Figure 1 supplies no controlled comparison, numerical accuracy or physical dynamics guarantee; the later tables must carry the empirical claim.
2. Motivation
2.1 The problem and the proposed response
The authors argue that future-image prediction entangles appearance with spatial dynamics, whereas metric point maps expose structure in a space aligned with motion planning. The practical question is whether predicting that geometry provides a better policy representation, beyond producing plausible future observations. e02
2.2 What this reading follows
A driving policy needs to anticipate how nearby space changes, not merely describe the current camera image. GeoWAM addresses this by predicting future geometric features and dense point maps, then using those features to infer ego motion. The key connection is asymmetric: the action decoder reads predicted geometry through a stop-gradient input. This edition follows that information flow from historical images to one future trajectory, then separates evidence about geometry accuracy from evidence about planning. The supplied v2 paper reports improvements on nuScenes and NAVSIM, while leaving the contribution of detachment, pretraining and future-token conditioning unablated. e01e04e07e08e11e12e14e16
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | IDM |
| Quadrant | Q4 · Dual-system × IDM |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded Q4 classification is supported by distinct geometry-forecasting and ego/action decoders connected through predicted future tokens. The paper explicitly calls this inverse-dynamics-like: motion is inferred from forecast scene evolution. Shared history and joint finetuning do not make it a single joint state/action generator. This is learned trajectory inference, not an exact physical inverse-dynamics solver. e04e07e08
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 Learn to forecast without giving the predictor future images
The geometry encoder first turns historical camera frames into a memory containing scene and ego information at several feature levels. Learned future queries retrieve that history and model evolution across future steps. Their features are projected into the representation expected by the shared point head, which predicts dense points in each future ego coordinate frame. Training can see future images, but only in a target branch: their geometric features are detached and compared with predictions using cosine alignment. Dense point supervision adds regression, confidence-aware and surface-normal terms, while current-frame geometry remains supervised. These are distinct constraints on feature similarity and spatial reconstruction. At inference the target branch's future images disappear; the forecast must come from history. Geometry pretraining precedes planning finetuning, which retains the same geometric objectives. e04e05e06e09
Figure 2. Future geometry conditions a distinct ego-motion decoder. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the historical frames through the image features and geometry encoder into the central memory. The legend separates orange ego/pose tokens from green geometry tokens. At the bottom, learned geometry queries attend to history and evolve across future time through temporal self-attention; Point DPT converts their predicted features into point maps. Follow the dashed green connection upward: future geometry also supplies keys and values to the pose decoder. Its orange output reaches the trajectory head. The caption and Equation (9) specify a stop-gradient connection, so the information passes forward while trajectory gradients are blocked through this particular future-geometry input. e04e05e07e08e09e16
What it supports. Geometry forecasting is used while producing the trajectory, rather than serving only as a training objective. The method couples separate forecast and action decoders through an explicit predicted-token interface. This supports the report's Dual-system × IDM assessment, with the paper's qualification that the formulation is inverse-dynamics-like.
Where the evidence stops. The diagram omits some trajectory-head refinement detailed on page 6. Detachment is not a globally frozen backbone: the branches share historical features and the full model is finetuned. No ablation establishes the benefit of that gradient boundary.
5.2 Read the one-way connection as an architectural constraint
The future ego decoder receives both historical memory and predicted geometry tokens. Equation (9) detaches the latter before cross-attention, expressing the authors' inverse-dynamics-like idea: infer ego motion from anticipated scene evolution. A trajectory head subsequently combines historical and predicted ego tokens and regresses one sequence of planar positions and headings. Reader interpretation: this makes future geometry a meaningful intermediate interface, but it does not prove the policy depends strongly on that interface, because historical memory is also available. Nor does detachment mean all geometry-related parameters are frozen; the paper jointly finetunes shared components. The reported comparisons do not remove either information path or alter the gradient boundary. Consequently, architectural compatibility with inverse dynamics is supported, while the causal importance of its future-conditioning path remains an experimental question. e07e08e09e16
5.3 Move from prediction accuracy to feedback-sensitive planning
The evidence has three separate levels. Table 1 asks whether predicted point maps preserve future ray depth on nuScenes validation; video baselines must first synthesize images and then reconstruct geometry. Table 2 asks whether the resulting policy achieves high NAVSIM navtest scores under the human-penalty protocol. Table 3 tests harder navhard scenes through original and synthetic reactive stages, where new rendered observations reflect the planned ego pose. Reader interpretation: the last setting is particularly relevant to accumulated planning errors, but it remains pseudo-simulation. GeoWAM leads the reported aggregate scores, yet its stage-two collision and lane-keeping components deteriorate substantially. Reading these tables together supports useful geometric prediction and competitive planning; it does not establish a matched-compute advantage, physical driving safety or which training component produced the gains. e10e11e12e13e14e16
5.4 Training and inference
During training
Geometry targets come from off-the-shelf geometry foundation models, without occupancy annotations. Future images enter a target branch of the same encoder only during training. Detached patch targets receive cosine alignment; point supervision combines Euclidean regression, confidence-aware regression and multi-scale surface-normal consistency. Current-frame point supervision anchors the encoder. e03e06
Pretraining combines OpenScene, nuScenes, Bench2Drive, Waymo Open Dataset, KITTI, Argoverse 2 and DDAD. The future decoder has six layers, width 1024 and 16 heads. Training predicts eight frames at 2 Hz from three history frames, dynamically sampling two to eight views. e10
DVGT-2 initializes the encoder and point head. Pretraining runs 161 epochs with AdamW, weight decay 0.05, bfloat16, 5% warmup and cosine decay. Peak learning rates are 10^{-4} for the future decoder and 2×10^{-5} for pretrained components. Planning finetunes the complete model for 40 navtrain epochs with eight views; the new action head also uses 10^{-4}. Both planning loss weights are 5. e04e10e09
During inference
Use history to forecast geometry tokens, decode future ego tokens and regress one trajectory; future ground-truth images are absent. Geometry prediction is an inference-time conditioning signal, while historical relative-pose supervision is auxiliary training. In navhard, benchmark-rendered observations from resulting ego poses feed subsequent planning calls; this is simulated feedback, not a demonstrated physical controller. e05e07e08e09e13
5.5 Implementation flow
- Encode history
DVGT-2 produces geometry tokens and ego tokens for each view, time and retained feature level. Their concatenation forms historical memory. Figure 2 depicts the image feature stage as DINOv3; the method identifies DVGT-2 as the geometry encoder. e04e10
- Forecast future geometry
Learned geometry queries combine a shared seed with time, view and 2D positional embeddings. Causal temporal self-attention links future steps at each spatial location; cross-attention retrieves history. Level-specific projections feed the shared Point DPT head, predicting one 3D point per pixel in each future ego frame and confidence. e05
- Infer ego motion from the forecast
Learned ego queries use causal temporal attention and cross-attend to both history and stop-gradient future geometry tokens. The action branch consumes latent geometry, rather than passing decoded point clouds into a separate geometric optimizer. Detachment blocks this conditioning path's trajectory gradients; it does not freeze the shared encoder. e07e04e09
- Regress a trajectory
A causal temporal transformer refines concatenated historical and predicted ego tokens. A learned trajectory query cross-attends to them, and a regression head outputs positions and headings. There are no trajectory anchors, mode classification or iterative sampling. e08
6. Experiments & results
GeoWAM forecasts dense future geometry from historical multiview images, then conditions a separate ego-motion decoder on predicted geometry tokens to regress one driving trajectory. Geometry pretraining and joint planning finetuning yield leading scores in the supplied NAVSIM tables, but the paper does not isolate which architectural component causes the gains.
The supplied 13-page v2 paper contains Figures 1–3 and Tables 1–3, all included here. Its experiment section reports geometry and planning baseline comparisons plus selected qualitative examples, but no controlled ablation or separate ablation page. Consequently no visual is labeled as an ablation; the horizon and stage breakdowns offer diagnostics but do not isolate a mechanism. The source also contains no physical deployment experiment. e11e14e15e16
6.1 Read the original evidence
Table 1. Geometry accuracy improves on average, with a short-horizon exception. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the two metric blocks separately: lower absolute relative ray-depth error is better, while higher threshold accuracy is better. Each block shows one-, two-, three- and four-second columns, followed by a mean over all eight predicted frames, not just the four displayed horizons. The first two rows forecast RGB with Epona or Cosmos 3 and then reconstruct geometry with DVGT. VGGT-World and GeoWAM forecast geometry directly. Section 4.1 defines ray depth as distance from the predicted 3D point to the ego origin and identifies nuScenes validation as the evaluation split. Bold and underlining mark the best and second-best entries. e10e11
What it supports. GeoWAM's mean Abs Rel is 0.257 versus 0.274 for Epona+DVGT, and its mean threshold accuracy is 0.754 versus 0.655. It has the lowest Abs Rel in every displayed horizon. The exception is one-second threshold accuracy: Epona+DVGT reaches 0.732, above GeoWAM's 0.708.
Where the evidence stops. These are end-to-end pipeline comparisons. Video baselines incur both video-generation and reconstruction errors, and the table does not establish matched pretraining data, compute or uncertainty. Ray-depth accuracy alone does not measure planning safety.
Table 2. The best navtest aggregate does not mean winning every component. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the rightmost EPDMS column to identify the overall ranking, then move left to inspect how each policy performs across driving components. Page 7 defines NC as no at-fault collision, DAC as drivable-area compliance, DDC as driving-direction compliance and TLC as traffic-light compliance. EP measures ego progress, TTC concerns time to collision, LK is lane keeping, and HC/EC are history/extended comfort. Higher is better under the official human-penalty protocol. Compare GeoWAM particularly with DVGT-2, its initialization, while remembering that the row comparison changes a complete planning system rather than one isolated component. e10e12e16
What it supports. GeoWAM scores 90.2 EPDMS, compared with 89.6 for DVGT-2 and 88.9 for EponaV2. It matches the best DDC and TLC entries at 99.7 and 99.9. Its ego progress is 87.0 versus DVGT-2's 87.9, while extended comfort is 86.8 versus 77.0, illustrating why the aggregate needs component-level context.
Where the evidence stops. Table 2 calls this closed-loop planning on navtest; it should remain distinct from the two-stage navhard protocol. The 0.6-point improvement has no reported error bars, and the table alone cannot attribute it to future geometry.
Table 3. navhard adds feedback, while retaining substantial stage-two degradation. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each method as a pair of rows: S1 evaluates original scenes, and S2 evaluates synthetic reactive scenes. The EPDMS column is merged across the pair and reports an overall method score; it is not a separate S2 score. The retained table note explains that gray dagger-marked methods use reinforcement learning or PDMS-score supervision, and that bold/underlined rankings exclude those rows. Section 4.3.2 describes the feedback mechanism: 3D Gaussian Splatting renders an observation from the resulting planned ego pose for the next planning step. Compare both the overall score and the changes in individual components between stages. e10e13e14
What it supports. GeoWAM reaches 36.6 overall EPDMS, above DVGT-2's 31.7 and the differently supervised EponaV2's 36.1. The stage breakdown is less reassuring: GeoWAM's NC falls from 97.7 to 80.4 and lane keeping from 96.0 to 49.9. Leading the aggregate therefore coexists with considerable difficulty in synthetic reactive scenes.
Where the evidence stops. This is pseudo-closed-loop evaluation, not physical driving. Differing baseline supervision matters, and the supplied paper does not give the complete aggregation formula; do not reconstruct EPDMS by averaging the visible components or stages.
Figure 3. Three selected maneuvers make the forecast representation visible. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the panels from left turn to straight driving to right turn, following their maneuver labels; the repeated '(a)' labels are present in the original. The colored boxes mark successive predicted ego poses, while each surrounding point cloud aggregates geometry from all future time steps. This is a temporal overlay, not a single predicted camera frame. Look at how the pose sequence passes through the reconstructed street layout and road markings. Section 4.4 draws attention to a following vehicle in the left-turn example and avoidance of a roadside vehicle in the straight-driving example. Those are the authors' interpretations of selected predictions. e15
What it supports. The visual illustrates how the method expresses an anticipated maneuver within a predicted geometric scene. Road structure and pose sequences remain recognizable in all three examples, supporting a qualitative claim of spatial coherence. It supplies an intuitive view of the representation that complements the numerical geometry and planning tables.
Where the evidence stops. No ground-truth sequence, calibrated error scale or failure frequency accompanies these panels. Aggregating time can obscure errors in individual frames, so the examples cannot establish accurate surrounding-agent dynamics or reliable obstacle avoidance.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Future geometry prediction on nuScenes nuScenes validation; ray depth measured from the ego origin; eight predicted frames over four seconds. | GeoWAM: 0.257; 0.754. Mean Abs Rel (lower better); mean threshold accuracy δ < 1.25 (higher better) | Epona+DVGT: 0.274; 0.655. VGGT-World: 0.325; 0.544. Cosmos 3+DVGT: 0.376; 0.503. Best reported means, but one-second threshold accuracy is lower than Epona+DVGT: 0.708 versus 0.732. Video baselines include reconstruction of generated RGB, so this is a pipeline comparison, not a controlled representation ablation. e10e11 |
| Planning on NAVSIM v2 navtest Finetuned on navtrain; navtest benchmark evaluation with the official human-penalty protocol. Table 2 labels this closed-loop planning. | 90.2. EPDMS (higher better) | DVGT-2: 89.6; EponaV2: 88.9; DriveLaW: 88.6. A 0.6-point gain over DVGT-2 and the highest aggregate score in this table. GeoWAM does not lead every component; no run-to-run uncertainty is provided. e10e12 |
| Two-stage planning on NAVSIM v2 navhard Original scenes in S1, synthetic reactive scenes in S2; 3D Gaussian Splatting pseudo-closed-loop feedback; official human-penalty protocol. | 36.6. Overall EPDMS (higher better) | DVGT-2: 31.7; DriveLaW: 30.6; EponaV2†: 36.1; NavFormer†: 34.1. Dagger denotes reinforcement-learning or PDMS-score supervision. Highest reported aggregate, including differently supervised baselines. The merged EPDMS cell summarizes the method, not S2 alone. GeoWAM's NC declines from 97.7 in S1 to 80.4 in S2; robustness remains limited. e10e13e14 |
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
The supplied experiments contain baseline comparisons and three selected qualitative cases, but no ablation of pretraining, loss terms, stop-gradient or future-token conditioning. Causal attribution to any one mechanism is therefore unresolved; there is no dedicated limitations section. e16
The geometry comparison includes video-generation and reconstruction errors, without a demonstrated matched-data or matched-compute design. NAVSIM results support benchmark planning, not physical road deployment; navtest and navhard scores must not be pooled. e11e12e13e14
Aggregated predicted point clouds illustrate spatial coherence but provide no per-frame ground-truth overlay, uncertainty or failure frequency. The deterministic head's behavior under multiple plausible futures is untested. e08e15e16
7.2 Questions for discussion
- Does future-token conditioning add value beyond the shared geometry encoder?
- Does preserving geometric accuracy through detachment improve recovery in synthetic reactive scenes?
8. Reproducibility audit
8.1 Requirements and known gaps
Required ingredients include DVGT-2 initialization, the seven-dataset pretraining mixture, point targets and masks, NAVSIM v2 splits and human-penalty scoring. The supplied text does not specify mixture weights, sample counts, batch size, image resolution, hardware, runtime or detailed target-generation/masking procedures. e03e06e10
Proposed checks: compare stop-gradient against an otherwise identical gradient-enabled model, measuring geometry and planning together; separately remove or temporally shuffle future geometry conditioning while preserving history and training budget. Both comparisons should use repeated seeds and stage-specific navhard metrics. e07e09e11e14
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 detaching future geometry preserve a useful forecast?
Reader-proposed experiment, not performed: initialize both variants from the same geometry-pretrained checkpoint and finetune with identical data, losses, learning rates and budgets. Change only Equation (9)'s stop-gradient operation. Across repeated seeds, measure nuScenes geometry before and after finetuning, navtest EPDMS and each navhard stage's components. The proposed preservation benefit is weakened if allowing gradients yields equal or better geometry and planning; a planning gain accompanied by geometric degradation would instead reveal a measurable tradeoff. e06e07e09e10e11e14
Check 2: Does the policy use the forecast beyond historical memory?
Reader-proposed experiment, not performed: compare the full model with a variant trained without future-geometry cross-attention, keeping historical memory, geometry losses and the action-head capacity and budget as closely matched as possible. For the full model, additionally shuffle future-token time order at evaluation and include an unchanged-input control. Report geometry metrics, navtest EPDMS and navhard S1/S2 components across seeds. No reproducible loss from removing the path would challenge its necessity; sensitivity to shuffling alone would show dependence on that input, not prove its causal superiority. e07e08e10e12e14e16
8.3 Reading coverage
Visual audit: Visually inspected the title/author/version block, all three figures, all three tables and the supporting methods, objectives, setup, feedback protocol and conclusion on pages 1–10. All six final original-PDF crops were inspected, including the navhard supervision/ranking note. Pages 11–13 contain references and were read as text. No appendix or controlled ablation is present in the supplied paper; separate supplements were not supplied.
PDF pages inspected for this edition: 1, 2, 3, 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, authors, abstract and version stamp (p. 1)
- 1 Introduction (pp. 1–3)
- 2 Related Work, including 2.1–2.3 (pp. 3–4)
- 3 Methodology: 3.1 Visual Geometry World Model and 3.2 GeoWAM (pp. 4–6), Eqs. (1)–(11)
- 4 Experiments: 4.1 Experimental Setup, 4.2 Future Geometry Prediction, 4.3 Planning on NAVSIM v2, 4.4 Visualization (pp. 6–9)
- 5 Conclusion (p. 10)
- References [1]–[49] (pp. 10–13)
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 limitation was addressed by inspecting PDF pages 1–10 and all three figures and tables. References on pages 11–13 were read as text.
- The supplied artifact is arXiv:2608.23486v2 [cs.CV], stamped 25 Aug 2026; its title block says August 26, 2026. The catalog submission date is August 25. Title and all 13 authors agree with the catalog after name-order normalization. Earlier editions were not supplied, so content changes between editions cannot be assessed.
- No separate supplement was supplied. No code was inspected and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author block, Date line and arXiv margin stamp
Exact title and 13 authors match the catalog. Affiliations are Uber AV Labs and Case Western Reserve University. Stamp: 2608.23486v2, 25 Aug 2026; internal date: August 26, 2026.
Go to primary source ↓e02PDF pp. 1–3, Section 1; p. 2, Figure 1
Authors motivate explicit point-cloud geometry as better aligned with driving dynamics than future pixels; Figure 1 contrasts the two prediction spaces.
Go to primary source ↓e03PDF p. 3, Section 2.1, final paragraph
Future geometry uses dense metric point-map targets derived from off-the-shelf geometry foundation models and does not require ground-truth occupancy annotations.
Go to primary source ↓e04PDF p. 4, Figure 2 and Sections 3–3.1; p. 5, Eq. (1)
Two-stage training shares a DVGT-2 historical geometry/ego memory between distinct forecasting and action branches; Figure 2 labels DINOv3 and Point DPT.
Go to primary source ↓e05PDF p. 5, Section 3.1, Future geometry decoding, Eqs. (2)–(4)
Learned position/time/view queries use causal temporal and historical cross-attention; projected future features yield point maps in future ego frames and confidence.
Go to primary source ↓e06PDF p. 5, Future geometry supervision, Eqs. (5)–(7)
Future-image targets are detached and excluded from inference. Losses combine cosine feature alignment and current/future point regression, confidence and surface normals.
Go to primary source ↓e07PDF p. 6, Section 3.2, Future ego-token decoding, Eqs. (8)–(9)
Ego queries attend to history and detached predicted future geometry; the stated inverse-dynamics-like design blocks trajectory gradients through future geometry conditioning.
Go to primary source ↓e08PDF p. 6, Trajectory decoding, Eq. (10)
Historical and future ego tokens feed a causal transformer and trajectory query; a single trajectory contains current-frame planar positions and headings without anchors or sampling.
Go to primary source ↓e09PDF p. 6, Planning objective, Eq. (11)
Finetuning retains geometric losses, adding L1 trajectory and auxiliary historical relative-pose losses.
Go to primary source ↓e10PDF p. 7, Section 4.1, Datasets, Metrics and Implementation details
Specifies seven pretraining datasets, NAVSIM splits, ray-depth metrics, human-penalty EPDMS components, 3-to-8-frame setup, decoder size, 161/40 epochs, learning rates and loss weights.
Go to primary source ↓e11PDF p. 7, Table 1, all rows and mean/1s columns; Section 4.2
Mean Abs Rel/threshold accuracy: GeoWAM 0.257/0.754, Epona+DVGT 0.274/0.655, Cosmos 3+DVGT 0.376/0.503, VGGT-World 0.325/0.544. Epona leads 1s threshold accuracy, 0.732 versus 0.708.
Go to primary source ↓e12PDF p. 8, Table 2, GeoWAM, DVGT-2, EponaV2 and DriveLaW rows; Section 4.3.1
navtest EPDMS is 90.2, 89.6, 88.9 and 88.6 respectively. GeoWAM DDC/TLC are 99.7/99.9; EP is 87.0 and EC 86.8.
Go to primary source ↓e13PDF p. 7, Section 4.1, Datasets; p. 8, Section 4.3.2
navhard uses original and synthetic reactive stages; 3DGS renders new observations from planned ego poses, allowing deviations to influence later predictions.
Go to primary source ↓e14PDF p. 9, Table 3, caption, EPDMS column and GeoWAM S1/S2 rows
Overall EPDMS: GeoWAM 36.6, DVGT-2 31.7, DriveLaW 30.6, EponaV2† 36.1, NavFormer† 34.1. Gray dagger rows use RL or score supervision. GeoWAM NC: 97.7/80.4; LK: 96.0/49.9.
Go to primary source ↓e15PDF p. 8, Section 4.4; p. 9, Figure 3 and caption
Three selected maneuver visualizations aggregate future geometry; boxes denote predicted ego poses. Authors describe preserved structures, a following vehicle and roadside avoidance.
Go to primary source ↓e16PDF pp. 7–10, Sections 4.1–4.4 and 5; Tables 1–3 and Figure 3
Empirical material consists of future-geometry and planning comparisons plus qualitative examples. No controlled ablation, error bars, physical deployment experiment or dedicated limitations section appears before the conclusion and references.
Go to primary source ↓8.5 Primary sources
GeoWAM: Visual Geometry World Action Models for Autonomous Driving ↗
PDF · 6,373 extracted words
Source fingerprint
f8f88d34a9321711e2717bd782eb6830f5b5ab2a32e20be09a59c58c442f2501