DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving
1. Paper overview
In one sentence: Predicting visual futures during training improves driving representations at large data scale, while an optional action expert reduces decoding cost without generating images during control. e-motivatione-wme-experte-scalinge-v2e-dynamic-failure
| At a glance | What to know |
|---|---|
| Research problem | Author claim The authors argue that supervising a large VLM with only low-dimensional waypoints leaves its capacity underused. They propose dense predictive image supervision to improve transferable driving representations. This supervision-deficit explanation is a hypothesis supported by comparisons, rather than a general impossibility result for action-only learning. e-motivatione-scaling |
| Core mechanism | Source description Two implementations accommodate discrete visual tokens and continuous visual features: autoregressive image-token prediction and conditional latent diffusion. e-wm |
| A key reported result | Large-data planning with visual supervision: VQ: 1.0563 m and 0.0392%; ViT collision: 0.0302%. ADE ↓ (m); collision ↓ (%). In-house 70M-frame training; 100 challenging test scenarios; base models without action experts. ADE uses a 3-second, six-waypoint trajectory. Action-only VQ: 1.4829 m and 0.0488%; action-only ViT collision: 0.0359%. Reported relative reductions include 28.8% VQ ADE and 15.9% ViT collision. At 700k, VQ ADE and ViT collision worsen, so gains are not uniform across settings. e-datae-scaling |
| Reading caution |
Core contributions
- Source description
Two implementations accommodate discrete visual tokens and continuous visual features: autoregressive image-token prediction and conditional latent diffusion. e-wm
- Source description
A 500M action expert couples to the large backbone through joint attention and permits controlled comparisons of three trajectory decoders. e-expert
Figure 2. Two visual prediction objectives train the driving representation. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the bottom, where language, images and previous actions enter their respective tokenization modules. The central backbone produces language, visual and action features. Follow the upper-left branch as next-token prediction of an image from its preceding context; Section 3.2 calls this the current image because time is indexed relative to that prefix. On the right, a frozen VAE encodes the next image, noise is added, and a trainable denoiser receives current visual/action features as conditioning. The downward green gradient marker shows supervision reaching those features. Snowflakes and flames distinguish frozen coding components from trainable components; they do not indicate inference-time routing. e-contexte-wm
What it supports. A discrete visual vocabulary permits the same autoregressive backbone to receive dense image-token supervision. Continuous ViT features instead receive that supervision through a diffusion branch. Both designs train predictive representations, and the paper explicitly bypasses image generation during driving. The illustration therefore explains the training signal, not an online search through generated futures.
Where the evidence stops. Read the causal-prefix definition in Equation (2) alongside the diagram: AR predicts V_t, whereas diffusion predicts I_{t+1}. Treating both as reconstruction of the fully observed current frame would erase the mechanism. Loss weights and detailed diffusion configuration remain unspecified.
2. Motivation
2.1 The problem and the proposed response
The authors argue that supervising a large VLM with only low-dimensional waypoints leaves its capacity underused. They propose dense predictive image supervision to improve transferable driving representations. This supervision-deficit explanation is a hypothesis supported by comparisons, rather than a general impossibility result for action-only learning. e-motivatione-scaling
2.2 What this reading follows
A driving VLA must compress a camera scene into a short trajectory, leaving most visual structure without direct supervision. DriveVLA-W0 adds an image-prediction task to that training problem. Its two implementations follow the backbone's visual representation: discrete tokens support autoregression, while continuous features condition a diffusion model. A smaller action expert then specializes trajectory generation. The central reading question is which evidence supports better representations, which supports faster decoding, and which merely shows plausible generated scenes. The paired scaling results are compelling at the largest setting, but intermediate regressions, comfort tradeoffs and documented failures constrain the broader claims. e-motivatione-wme-experte-scalinge-v2e-dynamic-failure
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 | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The deployed expert variant supports Dual-system through separate VLA/action parameter streams coupled by joint attention. Other mechanisms and Outside quadrants fit world modeling used as auxiliary training supervision, with no image rollout during control or inverse-dynamics action recovery. Qualification: the base VQ model shares visual/action autoregression within one backbone, so the architecture label does not describe every instantiation. e-wme-experte-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 Why the visual target depends on the information already visible
The two world-model branches solve different prediction problems because their conditioning differs. In the VQ branch, Equation (2) predicts the tokens of V_t from the prefix before that image plus previously generated image tokens. Calling V_t the current frame does not make this a reconstruction objective: the causal prefix has not already supplied the complete target image. In the ViT branch, the denoiser receives all current visual and action features together. Predicting that same observed image would allow a reconstruction shortcut, so the paper instead targets I_{t+1}. This explains the time-index difference visible in Figure 2. Reader interpretation: the important shared property is supervision about information beyond the available context, rather than a requirement that every backbone produce images with the same decoder. e-contexte-wm
Figure 3. Shared attention connects a large context model to three alternative action decoders. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read panel (a) upward: each expert applies its own normalization and feed-forward processing, while joint attention connects their token streams. Equation (4) specifies concatenation of queries, keys and values, followed by splitting the resulting outputs. Panels (b)–(d) then show alternative output procedures, not three decoders run together. Query-based decoding maps updated queries through an MLP to waypoints. The orange feedback arrows in the AR panel represent sequential token generation. The loop in the flow panel represents repeated continuous denoising before denormalization. Every option prefills the previous action, preserving a common temporal input across the comparison. e-experte-latencye-v1
What it supports. The expert architecture changes where repeated action-generation computation occurs while retaining access to the large model's context. The query expert produces a full trajectory in one pass; autoregression incurs token-dependent cost; flow matching incurs iterative cost. This makes decoder choice a joint accuracy and latency decision, rather than a uniform acceleration.
Where the evidence stops. The paper calls this MoE, but the shown mechanism is coupled transformer streams, without a depicted sparse routing gate. The 74.3 ms headline belongs to the query expert; it cannot be combined with the separate AR best-of-six score.
5.2 Separate representation learning from the cost of producing an action
The two-stage schedule makes the paper's efficiency claim easier to understand. First, 6VA pretraining uses both action and visual losses to shape the VLA representation. Second, the model processes 2VA sequences and learns from the action expert's output loss, while the VLA backbone remains trainable. This is not a frozen teacher supplying fixed features. Joint attention connects the large and small parameter streams, and the decoder determines how often action-generation computation repeats. Queries need one pass, AR needs sequential tokens, and flow matching follows an iterative continuous path. The measured 74.3 ms result applies to the query expert on H200. Reader interpretation: the design offers several operating points; the maximum benchmark score and minimum measured latency cannot be combined into one demonstrated configuration. e-traininge-experte-latencye-v1
5.3 What survives a stricter reading of the scaling argument
Table 3 gives the most direct evidence for the central training idea because it compares each base VLA with and without world modeling and excludes the action experts. Its large-data gains survive close reading, but its intermediate regressions rule out a uniform improvement claim. Table 4 asks a different question: which action decoder works best in each of two datasets? The ranking reversal is real in the table, yet data distribution and evaluation metrics change alongside scale. Finally, the appendix's FID comparison varies temporal context and contains a configuration inconsistency, so it cannot isolate image fidelity as the cause of better planning. Reader interpretation: the paper establishes useful empirical design choices under its reported protocols; causal explanations about supervision deficit, decoder capacity and physical understanding remain hypotheses for more controlled tests. e-scalinge-decodere-fidelitye-qualitative
5.4 Training and inference
During training
Stage one uses 6VA sequences and action plus weighted world-model loss. Stage two switches to 2VA and supervises only the action expert output; the backbone remains trainable. NAVSIM experiments use 8k NuPlan pretraining steps followed by 4k fine-tuning steps; VQ is the default ablation model. e-training
NAVSIM training uses 256×144 images, eight NVIDIA L20 GPUs, global batch 48, AdamW, cosine scheduling, initial learning rate 2×10^-4 and bfloat16. In-house training uses 50k pretraining plus 30k fine-tuning steps on the same data, 64 GPUs and batch 256. e-training
During inference
Driving bypasses both AR visual generation and diffusion sampling. The action expert outputs waypoints by one query pass, sequential token generation, or iterative flow integration. Observations and past actions provide context; no online image-rollout search or low-level vehicle controller is specified. e-wme-contexte-experte-latency
5.5 Implementation flow
- Build the causal multimodal context
Interleave language, visual representations and previous-action tokens. Emu3 8B handles VQ images; Qwen2.5-VL 7B handles ViT features. Final hidden states are separated by modality. Baseline action decoding predicts FAST tokens and detokenizes them into waypoints. e-context
- Learn predictive representations
The VQ model predicts the current image token sequence from its preceding prefix; it is future prediction relative to that prefix. The ViT model denoises the next-frame latent using current visual/action features. Figure 2 marks frozen coding modules and a trainable backbone/denoiser. e-wm
- Specialize trajectory decoding
The large VLA and smaller action expert concatenate queries, keys and values for joint attention, then route outputs back to their respective streams. All experts prefill previous actions. Query decoding regresses waypoints with L1 loss; AR uses cross-entropy; flow matching fits a conditional vector field with MSE. e-expert
6. Experiments & results
DriveVLA-W0 trains driving VLAs to predict visual futures as well as actions, then uses an optional small action expert for efficient trajectory generation. Its strongest controlled evidence is improved large-data planning, while its visual simulator remains a diagnostic capability. Driving inference skips image generation (e-wm, e-scaling, e-expert).
6.1 Read the original evidence
Table 3. Paired comparisons show strong large-data gains and mixed intermediate-scale effects. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare each VLA baseline with the immediately following '+ World Model' row within the same data-scale block. The table excludes action experts, making these pairs more informative about visual supervision than the separately optimized benchmark scores. ADE is in meters over the in-house three-second trajectory; collision is in percent and lower is better for both. Green upward annotations denote relative improvements even though the metric values decrease; red downward annotations denote deterioration. Read the absolute entries first, then their annotations. The TransFuser rows provide additional baselines, but change architecture and initialization rather than isolating the world-model loss. e-scalinge-datae-baselinese-scale-conflict
What it supports. At 70M frames, VQ ADE falls from 1.4829 to 1.0563 m, a reported 28.8% reduction; ViT collision falls from 0.0359% to 0.0302%, a reported 15.9% reduction. At 700k, VQ ADE and ViT collision instead worsen. The strongest supported result is therefore a large-data benefit under this protocol, with exceptions at smaller settings.
Where the evidence stops. Figure 1 labels its scales 700K/7M/70M and annotates 20.4%, unlike the 70k/700k/70M table and main text. This discrepancy remains unresolved. Three scales and no reported uncertainty do not establish a universal power law or a guaranteed benefit from extra data.
Table 2. The aggregate benchmark lead includes a pronounced comfort tradeoff. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the shaded rightmost EPDMS column, then move left to see which criteria contribute to the aggregate. NC means no at-fault collision; DAC, drivable-area compliance; DDC, driving-direction compliance; and TLC, traffic-light compliance. The remaining columns are ego progress, time to collision, lane keeping, history comfort and extended comfort. Section 4.1 combines compliance penalties with weighted subscores, so a leading aggregate need not mean every component improves. Compare the final row with DiffusionDrive immediately above it. Keep the dash in ARTEMIS's extended-comfort cell as missing information; it is not a zero or a measured failure. e-v2e-datae-v1
What it supports. DriveVLA-W0 reports 86.1 EPDMS against DiffusionDrive's 84.5, a 1.6-point advantage in this table. However, extended comfort is 58.9 versus 87.7. Reading the component columns therefore changes the conclusion from blanket superiority to a specific aggregate planning gain with a visible weakness.
Where the evidence stops. Table 2 does not identify the action-expert or candidate-selection variant for its DriveVLA-W0 row. Do not import that setting from the differently annotated NAVSIM v1 table, or treat these aggregate benchmark scores as physical deployment evidence.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Large-data planning with visual supervision In-house 70M-frame training; 100 challenging test scenarios; base models without action experts. ADE uses a 3-second, six-waypoint trajectory. | VQ: 1.0563 m and 0.0392%; ViT collision: 0.0302%. ADE ↓ (m); collision ↓ (%) | Action-only VQ: 1.4829 m and 0.0488%; action-only ViT collision: 0.0359%. Reported relative reductions include 28.8% VQ ADE and 15.9% ViT collision. At 700k, VQ ADE and ViT collision worsen, so gains are not uniform across settings. e-datae-scaling |
| NAVSIM v1 trajectory planning Published NAVSIM comparison; DriveVLA-W0 uses one front-view camera. | 88.4 query; 90.2 query with multiple anchors; 93.0 AR best-of-six. PDMS ↑ | Listed AutoVLA dagger configuration: 92.1, three cameras. The peak score depends on candidate selection and cannot be assigned to the plain AR expert. Sensors and protocols differ across methods. e-datae-v1 |
| NAVSIM v2 trajectory planning Published NAVSIM v2 comparison; exact DriveVLA-W0 decoder/selection variant is not identified in Table 2. | 86.1 EPDMS; 58.9 extended comfort. EPDMS ↑; extended comfort ↑ | DiffusionDrive: 84.5 EPDMS; 87.7 extended comfort. A 1.6-point aggregate advantage coexists with much worse extended comfort; this is not improvement on every driving criterion. e-datae-v2 |
| Action-decoder comparison across datasets Same pretrained initialization across experts within each dataset; NAVSIM 103k versus in-house 70M frames. | Query/flow/AR: 88.4/87.2/85.3 PDMS. In-house AR: 1.0069 m, 0.0295%. PDMS ↑; ADE ↓ (m); collision ↓ (%) | In-house query: 1.1248 m, 0.0453%; flow: 1.0362 m, 0.0398%. The ranking reverses, but dataset, trajectory distribution and metric also change. Scale alone is not isolated. e-decodere-data |
| Transfer and sequence design NuPlan pretraining followed by NAVSIM fine-tuning; world-model ablations exclude experts. | VQ world-model transfer: 80.7 to 85.6; 6V versus 6VA pretraining: 84.1 versus 85.6. PDMS ↑ | Action-only VQ transfer: 68.7 to 62.2. Supports useful predictive pretraining and action interleaving. One-second two-frame input scores 85.6 versus 84.3 for four seconds; longer temporal gaps are not universally better. e-transfere-sequencee-horizon |
| Action-expert latency H200 measurement; query-based expert versus backbone-only DriveVLA-W0. | 74.3 ms and 88.4 PDMS. Latency ↓ (ms); PDMS ↑ | 117.8 ms and 85.6 PDMS without expert. 63.1% of original latency means a 36.9% reduction. AR latency depends on token length; the best-of-six headline score is not this latency configuration. e-latencye-v1 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Tables 5 and 6. Action interleaving and temporal context help, but the controls answer different questions. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The left table is Table 5: read pretraining down the first column while checking that fine-tuning remains 2VA. The slash denotes no pretraining, 6V denotes the vision-only configuration, and 6VA adds interleaved actions. The right table is Table 6: read both sequence columns together, because the VA row also changes fine-tuning. NC and DAC measure collision and drivable-area compliance; shaded PDMS summarizes planning. These experiments use the default VQ model without action experts. For the cleaner pretraining-context comparison on the right, compare 2VA with 6VA, since both retain 2VA fine-tuning. e-sequencee-traininge-fidelity
What it supports. With fixed 2VA fine-tuning, PDMS progresses from 80.7 without pretraining to 84.1 after 6V and 85.6 after 6VA. In Table 6, extending pretraining from 2VA to 6VA improves PDMS from 84.2 to 85.6. Both comparisons support richer pretraining context under the reported setup.
Where the evidence stops. The VA-to-6VA comparison changes both pretraining and fine-tuning context, while 6V-to-6VA changes action availability and supervision. These controls support the design but do not prove causal dynamics learning. Appendix Table 8 also leaves a sequence-label/score inconsistency unresolved.
Figure 12. A plausible scene can still omit a vehicle that matters to the plan. Original paper, p. 20 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the two input frames first, then move left to right through the four future columns. Compare the ground-truth row with the row conditioned on the ground-truth trajectory: the red circles draw attention to an oncoming vehicle that the generated future misses. This comparison holds the supplied trajectory condition to the recorded trajectory, so the highlighted visual error is not simply a consequence of conditioning on the planner's different turn. The bottom row instead uses the predicted trajectory. On the map, this figure's legend assigns green dots to ground truth and red dots to the prediction. The caption states that all displayed images are MoVQGAN-token reconstructions. e-dynamic-failuree-wm
What it supports. The example exposes a task-relevant weakness that broad scene realism can conceal: the generated intersection remains recognizable while a conflicting vehicle is absent. The authors associate that prediction error with the incorrect left-turn plan. It motivates object-level diagnostics alongside aggregate planning scores and image-quality measures.
Where the evidence stops. This is a selected diagnostic case, not a measured failure frequency. The caption's causal narrative does not mean generated images are consumed online: Section 3.2 says driving bypasses visual generation. The image therefore illustrates a representation weakness, not a documented rollout-based control chain.
7. Analysis & limitations
7.1 What the evidence leaves open
No uncertainty estimates accompany the main comparisons. Three data scales do not establish a universal power law or prove action-only scaling can never close the gap. The 100-scenario test description omits the collision denominator and precise split identifiers. e-scalinge-data
Figure 1 uses 700K/7M/70M and a 20.4% annotation; Table 3 and the main scaling text use 70K/700K/70M. The report retains Table 3 values without reconciling the unexplained discrepancy. e-scale-conflict
Table 8 associates better FID with better PDMS, but changes context length and inconsistently references Table 5; its 2VA score also differs from Table 6. This does not isolate visual fidelity as the causal driver. e-fidelity
The authors show failures from ambiguous navigation at a Y-junction and missed oncoming vehicles. Counterfactual image examples do not establish calibrated dynamics or physical driving safety; the failure images are token reconstructions. e-command-failuree-dynamic-failuree-qualitative
7.2 Questions for discussion
- Would visual supervision retain its advantage at equal total training compute?
- Does the decoder ranking reverse within one fixed dataset as only its training size changes?
- Which uncertainty or object-level diagnostic would detect the oncoming-vehicle failure before committing to a turn?
8. Reproducibility audit
8.1 Requirements and known gaps
Required components include the chosen pretrained backbone, visual coding modules, FAST, NuPlan/NAVSIM and the two-stage schedule. Reproducing the scaling result additionally requires the in-house data. Loss weights alpha/beta, detailed diffusion configuration, split identifiers and software versions remain unspecified; Appendix D mainly expands the TransFuser baseline. e-contexte-wme-traininge-datae-baselines
Proposed checks should hold data and compute fixed when testing visual supervision, and vary data scale within one dataset when comparing decoders. Record uncertainty, action-token lengths and matched inference budgets before attributing gains to causal dynamics or decoder capacity. e-scalinge-decodere-latency
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 aligned action context improve prediction beyond visual pretraining?
Reader-proposed check, not performed: repeat the Table 5 comparison with the same VQ initialization, NuPlan frame subsets, 2VA NAVSIM fine-tuning, seeds and optimizer schedule. Compare action-only pretraining, 6V and aligned 6VA, reporting both update-matched and total-compute-matched results. At visual-prediction evaluation, additionally shuffle only the conditioning action histories among comparable clips and measure the change in future-image quality and oncoming-vehicle recall. Keep downstream planning evaluation unshuffled and report PDMS with uncertainty. An aligned-6VA planning benefit that survives compute matching, together with sensitivity to corrupted action context, would support the proposed mechanism. No benefit, or no sensitivity to action corruption, would weaken that explanation. e-contexte-wme-traininge-sequencee-scalinge-dynamic-failure
Check 2: Does decoder ranking reverse when only training-data size changes?
Reader-proposed check, not performed: form nested training subsets from one fixed driving corpus and keep its test scenarios and trajectory horizon unchanged. At each size, initialize query, AR and flow-matching experts from the same pretrained checkpoint, as in Table 4. Compare a fixed training budget and a separate convergence-controlled budget, then evaluate ADE, collision and latency on the same hardware. Record AR token-length distributions and compare both single-output and explicitly matched candidate-selection budgets; sweep the flow step count rather than assuming ten steps is optimal. A reproducible AR crossover within this fixed distribution would support a scale-dependent advantage. A stable ranking, or a crossover removed by additional flow training, would challenge the scale-only explanation. e-decodere-datae-experte-latencye-v1
8.3 Reading coverage
Visual audit: The title/author block, Figures 1–14, Tables 1–9, equations, training/evaluation text and Appendix B–D details were visually inspected on these pages. All six final original crops were separately inspected at their native rendered resolution. Figure 2's frozen/trainable markers, conditioning and gradient direction were checked against Section 3.2 and Eqs. (2)–(3); Figure 3's feedback loops were checked against Section 3.3 and Eq. (4). No claim-relevant architecture-marker conflict was found. Figure 1's scale labels conflict with Table 3; Appendix B.2 has the disclosed Table 5/6/8 configuration inconsistency. Reference-only pages 11–14 were read in the complete text but are not part of the visual audit. Separate supplements, external code, data and deployment were not inspected. The failure crop is a qualitative diagnostic, not a controlled ablation.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 16, 17, 18, 19, 20, 21, 22. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract; Sections 1–2: introduction and related work (pp. 1–3)
- Sections 3.1–3.3: baseline, world modeling, action expert (pp. 3–6)
- Sections 4.1–4.6: data, implementation, comparisons, scaling, ablations, visualization (pp. 6–10)
- Sections 5–6: conclusion and acknowledgments; references (pp. 10–14)
- Appendix A; B.1–B.3; C.1–C.4; D–F (pp. 15–22)
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.
- All seven supplied text chunks were read individually, including references and appendices. No separate supplement was supplied; separate supplemental material availability has not been fully verified.
- The supplied extraction does not reconstruct figure images. This limitation was addressed by inspecting original PDF figures and tables on pages 1–10 and 15–22; reference-only pages 11–14 were read as text.
- Identity: the inspected title and all 13 authors agree with the catalog, apart from capitalization and extraction spacing. The artifact is labeled an ICLR 2026 conference paper. No numbered revision or revision history is supplied; differences from earlier submissions cannot be established. The supplied observedTitle is preserved verbatim in source metadata.
- Code, datasets and external links were not inspected; no installation or experiment was reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title and author block, conference header
The supplied work is the ICLR 2026 conference paper DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving. All 13 catalog authors appear, with Lu Hou in title case. The affiliations are CASIA and Yinwang Intelligent Technology Co. Ltd. No numbered revision is printed.
Go to primary source ↓e-motivationPDF pp. 1–3, Abstract, Section 1 and Section 2, World Models in Driving and Robotics
The authors identify sparse waypoint supervision as a bottleneck and propose predictive image supervision for representation learning, distinguishing this use from world models used primarily as data synthesizers.
Go to primary source ↓e-contextPDF pp. 3–4, Section 3.1, input sequence and action prediction; Eq. (1)
Language instructions, front-view images and past FAST-tokenized actions form an interleaved history. Emu3 8B handles VQ tokens and Qwen2.5-VL 7B handles continuous visual features under causal attention. Generated action tokens are detokenized into waypoints.
Go to primary source ↓e-wmPDF pp. 4–5, Figure 2 and Section 3.2, Eqs. (2)–(3), both Inference paragraphs
The AR objective predicts visual tokens from their preceding prefix. The diffusion objective predicts noise in the future-image latent conditioned on current visual and action features. Weighted visual and action losses train the backbone; explicit image generation is bypassed for driving inference. Figure 2 marks frozen tokenization/visual coding modules and trainable backbone/denoiser.
Go to primary source ↓e-expertPDF pp. 5–6, Section 3.3, Eq. (4), Figure 3 and decoder descriptions
A 500M action expert and the large VLA expert concatenate their Q/K/V token sequences for joint attention, then split outputs. Previous-action prefilling is shared across query regression, autoregressive token prediction and continuous flow matching. Their losses are L1, cross-entropy and vector-field MSE respectively.
Go to primary source ↓e-trainingPDF p. 7, Section 4.2, Two-stage Training Paradigm, NAVSIM and In-house Dataset
Stage one uses 6VA with world-model and action losses; stage two uses 2VA with only the action expert loss, leaving the backbone trainable. NAVSIM uses NuPlan pretraining for 8k steps and 4k fine-tuning steps, 256×144 images, eight L20 GPUs, batch 48, AdamW, cosine schedule, initial learning rate 2e-4 and bfloat16. In-house training uses 50k plus 30k steps, 64 GPUs and batch 256; the GPU model is not stated for this cluster.
Go to primary source ↓e-dataPDF p. 6, Section 4.1, NAVSIM v1/v2 metrics and In-house Dataset
PDMS and EPDMS are composite planning scores. The in-house training set contains 70 million frames from over one million clips; its test set contains 100 challenging scenarios. ADE is measured on a 3-second, six-waypoint, 2 Hz trajectory. Collision rate follows the NAVSIM no-at-fault-collision methodology; its evaluation denominator is not detailed here.
Go to primary source ↓e-baselinesPDF p. 8, Re-implemented TransFuser; pp. 21–22, Appendix D
The controlled TransFuser variants use one front-view camera, ResNet-34 or pretrained DINOv3 ViT-7B, and latent BEV queries in place of LiDAR features. Multi-scale cross-attention fuses camera and BEV latent representations for waypoint prediction.
Go to primary source ↓e-v1PDF p. 7, Table 1, DriveVLA-W0 rows and caption symbols; AutoVLA dagger row
DriveVLA-W0 scores 88.4 PDMS with query decoding, 90.2 with multiple trajectory anchors, and 93.0 with AR best-of-six. The listed AutoVLA dagger row scores 92.1 using three cameras, while DriveVLA-W0 uses one. These are distinct decoding protocols.
Go to primary source ↓e-v2PDF p. 7, Table 2, DriveVLA-W0 and DiffusionDrive rows, EPDMS and EC columns
DriveVLA-W0 reports 86.1 EPDMS versus DiffusionDrive 84.5. Their extended-comfort values are 58.9 and 87.7 respectively. The DriveVLA-W0 row has no action-expert or selection-protocol symbol.
Go to primary source ↓e-scalingPDF pp. 8–9, Section 4.4 and Table 3, VQ and ViT baseline/world-model rows
Scaling experiments exclude action experts. At 70M frames VQ ADE changes from 1.4829 to 1.0563 m and collision from 0.0488% to 0.0392%; ViT collision changes from 0.0359% to 0.0302%. At 700k, VQ ADE worsens from 1.5424 to 1.5985 m and ViT collision worsens from 0.0462% to 0.0513%.
Go to primary source ↓e-scale-conflictPDF p. 1, Figure 1(b); p. 8, Section 4.4 data-scale sentence; p. 9, Table 3 headings
Figure 1(b) labels 700K, 7M and 70M frames with collision in per-ten-thousand units and a 20.4% annotation. Section 4.4 and Table 3 instead specify 70k, 700k and 70M, with collision in percent. The source does not reconcile the scale settings or headline annotation.
Go to primary source ↓e-decoderPDF p. 9, Table 4 and preceding Action experts reverse performance with data scaling paragraph
Identical pretrained initialization is used across the three experts within each dataset. NAVSIM query, flow and AR experts score 88.4, 87.2 and 85.3 PDMS. At 70M in-house frames they report ADE 1.1248, 1.0362 and 1.0069 m, and collision 0.0453%, 0.0398% and 0.0295%.
Go to primary source ↓e-sequencePDF pp. 9–10, Section 4.5 and Tables 5–6
The world-model ablations exclude action experts. Table 5 keeps 2VA fine-tuning fixed: no pretraining, 6V pretraining and 6VA pretraining score 80.7, 84.1 and 85.6 PDMS. Table 6 scores VA/VA, 2VA/2VA and 6VA/2VA at 83.3, 84.2 and 85.6.
Go to primary source ↓e-latencyPDF p. 10, Ablation on Latency; pp. 15–16, Appendix B.1 and Figure 6
The H200 query-expert comparison reports 117.8 to 74.3 ms and PDMS 85.6 to 88.4. Appendix B.1 uses ten flow-matching steps, approximately 145 ms, versus one query pass. AR latency varies with token length: about 95 ms at NAVSIM average 5.6 tokens and 170 ms at in-house average 17.8 tokens.
Go to primary source ↓e-transferPDF p. 8, Figure 4 and Section 4.4; p. 15, Table 7
NuPlan pretraining followed by NAVSIM fine-tuning changes VQ world-model PDMS from 80.7 to 85.6, and action-only VQ from 68.7 to 62.2. ViT action-only improves slightly, 70.3 to 70.6, so detrimental transfer does not apply to every baseline.
Go to primary source ↓e-fidelityPDF p. 16, Appendix B.2 and Table 8; p. 10, Tables 5–6
Table 8 pairs 2VA and 6VA with FID 9.847 and 4.610 and PDMS 84.1 and 85.6; MoVQGAN reconstruction FID is 3.007. Its prose references the bottom two rows of Table 5, which actually compare 6V and 6VA. Table 6 gives 84.2 for 2VA/2VA, leaving a configuration/score inconsistency.
Go to primary source ↓e-horizonPDF p. 17, Appendix B.3 and Table 9
From one 6VA checkpoint, current-frame-only fine-tuning yields 82.9 PDMS, two frames four seconds apart yield 84.3, and two frames one second apart yield 85.6. This is an input-interval comparison, not a measured long-horizon rollout benchmark.
Go to primary source ↓e-qualitativePDF p. 10, Section 4.6 and Figure 5; p. 20, Figure 13; p. 21, Appendix C.4 and Figure 14
Generated examples show different futures under ground-truth, stopping/decelerating and counterfactual right-turn actions. Figure 13 presents plausible intersection and traffic-jam predictions. These displays supply qualitative examples, without quantitative counterfactual accuracy or physical-driving validation.
Go to primary source ↓e-command-failurePDF pp. 18–19, Appendix C.2 and Figure 11 caption
A generic go-straight command at a Y-junction is ambiguous; the predicted trajectory enters the fork. The caption identifies the coarse navigation command set as a limitation and says displayed images are reconstructed from MoVQGAN tokens, not raw RGB frames.
Go to primary source ↓e-dynamic-failurePDF p. 20, Figure 12 and caption
The model misses an oncoming vehicle in the future conditioned on the ground-truth trajectory; the authors associate this with an incorrect left-turn plan. The displayed frames are MoVQGAN-token reconstructions. Green and red dots in this figure identify ground-truth and predicted trajectories respectively.
Go to primary source ↓e-trajectory-visualsPDF pp. 17–19, Appendix C.1 and Figures 7–10
Selected trajectory diagrams compare world modeling with baselines and depict unstable flow-matching trajectories versus more stable AR trajectories. These are illustrative cases, not a separate aggregate stability metric.
Go to primary source ↓8.5 Primary sources
DRIVEVLA-W0: WORLDMODELSAMPLIFYDATA SCALINGLAW INAUTONOMOUSDRIVING ↗
PDF · 10,268 extracted words
Source fingerprint
1a29df41a78a27e0a9f6e9a36580f2610a336c28dea79821cef984a563f579d5