A Survey on Vision-Language-Action Models for Autonomous Driving
1. Paper overview
In one sentence: The survey organizes driving VLAs by where language enters action generation, while exposing the need to evaluate instruction following and executed driving under compatible protocols. e-gape-architecturee-stagese-evaluatione-inconsistencies
| At a glance | What to know |
|---|---|
| Research problem | Author claim Driving requires perception, interpretation of traffic context and timely control. The authors argue that language-capable perception leaves an action gap when explanations do not reliably govern the vehicle. They frame VLA4AD as integrating language with action generation, while acknowledging that fluent rationales alone provide no formal safety guarantee. The survey aims to organize this fragmented literature. e-gap |
| Core mechanism | Source description An architectural vocabulary separates multimodal sensing, language processing and action decoding. A historical taxonomy distinguishes explainers, modular VLAs, unified policies and reasoning augmentation. These are organizing abstractions across heterogeneous systems, rather than specifications of one implementation. e-architecturee-stages |
| Reading caution | Reader analysis No new controlled results or ablations are reported. Secondary claims include TS-VLM BLEU-4 56 and a DiffVLA error-halving statement, but the survey lacks sufficient splits, baseline values, metric details or uncertainty for a reproducible comparison. They are not promoted to structured results. Dataset counts describe inventory scale, not driving performance. e-evidence-boundarye-datasets |
Core contributions
- Source description
An architectural vocabulary separates multimodal sensing, language processing and action decoding. A historical taxonomy distinguishes explainers, modular VLAs, unified policies and reasoning augmentation. These are organizing abstractions across heterogeneous systems, rather than specifications of one implementation. e-architecturee-stages
- Source description
Table 1 compares 15 representative model rows by modalities, data, components and outputs. Table 2 inventories eight dataset/benchmark rows. The abstract claims coverage of over twenty models across the review; that should not be confused with Table 1 alone or interpreted as a quantitative meta-analysis. e-gape-model-tablee-datasets
- Author claim
The authors connect training practice to three evaluation axes: executed closed-loop driving, open-loop prediction and robustness stress. They recommend measuring control, language and their interaction together, because current suites often separate them. e-traininge-evaluation
Figure 1. Follow the action-producing branch to understand the survey’s motivating distinction. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with panel (a): sensor and environment information flows through an end-to-end network into driving actions. Panel (b) sends sensor input to a VLM, then downward to reasoning chains and onward to multiple tasks. The Database arrow points left into the VLM and is labeled Finetune; it should not be read as an inference-time retrieval loop. In panel (c), follow the arrows from sensor input through the multimodal encoder to the LLM/VLM, downward into the action decoder, and leftward to driving actions. The conceptual difference is the explicit route from language-capable processing to action generation, not merely the presence of a language model. e-gape-architecturee-evaluation
What it supports. The figure explains the survey’s action-gap argument: language reasoning becomes relevant to a driving policy when it participates in producing the plan or controls. It supports a comparison of information pathways. It does not quantify how much safety, interpretability or generalization any pathway provides.
Where the evidence stops. The caption describes robust closed-loop control, but the drawing contains no environment feedback arrow or supporting performance measurement. Treat this as the authors’ architectural motivation; Sections 3.3 and 6.2 supply the execution and evaluation distinctions.
2. Motivation
2.1 The problem and the proposed response
Driving requires perception, interpretation of traffic context and timely control. The authors argue that language-capable perception leaves an action gap when explanations do not reliably govern the vehicle. They frame VLA4AD as integrating language with action generation, while acknowledging that fluent rationales alone provide no formal safety guarantee. The survey aims to organize this fragmented literature. e-gap
2.2 What this reading follows
A vehicle can describe a pedestrian and still fail to yield. This survey uses that gap between understanding and action to organize vision-language-action research for autonomous driving. Its figures progress from conventional sensor-to-control networks to explicit action decoders and reasoning-augmented policies; its tables inventory the models and data behind those categories. Read these visuals as a vocabulary for asking architectural questions. They do not establish a performance hierarchy. The useful endpoint is a sharper evaluation question: does language change the vehicle’s behavior appropriately while preserving safety? Internal inconsistencies in several model summaries make the boundary between survey description and independently verified implementation especially important. e-gape-architecturee-stagese-evaluatione-inconsistencies
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 Related resources. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The foundational survey/technical-resource placement is supported. Architecture, prediction paradigm and quadrant are correctly not applicable to this survey: it reviews modular and unified systems and different future/action mechanisms without proposing one model. Joint-training examples and conceptual boxes do not establish a One Model classification for the individual systems. e-architecturee-stagese-model-table
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 |
|---|---|
|
|
5. Method in detail
5.1 1. Trace the command all the way to the vehicle
The survey’s central question is operational: where does language enter the action-producing path? Begin with Figure 1’s separation between reasoning outputs and a decoder that emits driving actions. Figure 3 then divides that route into modular and unified alternatives. In the modular case, an intermediate representation is consumed by an action head; in the unified case, the representation and action prediction belong to one multimodal pathway. Section 3.3 adds a necessary distinction that the cartoons compress: a waypoint sequence may still need an MPC controller to execute it. Reader analysis: to assess a particular system, identify the command input, the predicted object, the controller and the source of fresh observations. A shared diagram or fluent explanation cannot substitute for that complete account. e-gape-architecturee-stages
Figure 2. The shared components describe an interface, not one fixed VLA implementation. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the large blocks from left to right: Multimodal Input, LLMs, and Action Decoder. The smaller labels illustrate information or capabilities associated with each stage: detection and occupancy at the visual side, environment description and future prediction near language processing, and trajectory planning or vehicle control at the output. Section 3 supplies the details omitted by the cartoon: cameras may be combined with vehicle state and other sensors; instructions and queries provide language input; decoders may generate sequential tokens or continuous trajectories. The separate command input is described in the text rather than drawn as a distinct arrow. The Qwen logo is illustrative, not a universal backbone requirement. e-architecturee-worldmodels
What it supports. The survey’s architectural abstraction separates representation, language processing and action decoding, while retaining alternative output types. A predicted trajectory can require an MPC controller to execute it. The Future Prediction label alone does not demonstrate an action-conditioned world model, a particular training loss or inference-time simulation.
Where the evidence stops. The paper supplies no tensor shapes, attention mask, shared-parameter map or loss equation for this diagram. Its components and arrows cannot establish a One Model classification, and steering/braking icons do not document a deployed control loop.
5.2 2. Separate what is learned from what runs online
Section 6.1 describes a family of training practices: imitate expert image–text–action sequences, align pretrained components, add targeted data or fine-tuning, and adapt the resulting policy for deployment. These stages do not imply that every component is trained together or that reasoning text receives the same supervision as actions. The survey explicitly notes that joint text/control optimization remains rare. During inference, its examples range from modular intermediate plans to unified decoding and reasoning with memory. The ADriver-I discussion introduces another mechanism, action-conditioned future-frame prediction, which should be distinguished from ORION’s memory-based reasoning. Reader analysis: classify a world-model contribution by the prediction target and its actual use in choosing actions, not by the words future, reasoning or VLA in an illustration. e-traininge-stagese-worldmodels
Figure 3. Compare the role of the intermediate representation before treating these as historical stages. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each column is a top-to-bottom information path. The first ends in explanation, question answering and description. The second introduces an explicitly labeled Intermediate Representation between the VLM and Action Head, which then produces trajectory or control outputs. The third collapses the action-producing route into an end-to-end VLM pathway. The fourth replaces the VLM block with Reasoning VLMs and Tool-use Agents while preserving an action output. The caption and Section 4 describe these as stages of development. Reader interpretation: the most useful comparison is whether language is descriptive, mediates a separate planner, or participates in a unified action pathway; chronological placement alone does not establish implementation details. e-stagese-model-tablee-inconsistencies
What it supports. The taxonomy makes an architectural tradeoff visible: explicit intermediate plans can aid inspection but introduce interfaces, whereas unified and augmented pathways couple language and action more directly. Section 4 attributes latency and potential failures to modular interfaces. The figure contains no controlled comparison proving that later stages outperform earlier ones.
Where the evidence stops. The categorical boundaries are imperfect in the source: Section 4.1 names DriveGPT-4 as an explainer, while Table 1 assigns it low-level control output. Preserve that tension instead of inferring that every named example matches its column exactly.
5.3 3. Read the tables as a map of evidence requirements
Table 1 tells the reader which interfaces and data a model is said to use; Table 2 tells the reader what different resources can test. Neither is a leaderboard. Section 6.2 makes the next step explicit: distinguish open-loop trajectory prediction, closed-loop route completion and infractions, and language or sensor stress. Reader analysis: a persuasive instruction-following result needs both appropriate behavioral change and acceptable driving outcomes under a documented protocol. The survey does not supply that common comparison, and some of its descriptions conflict. For example, its DriveGPT-4 explainer placement and low-level-control table entry point to a category boundary needing verification. Use these discrepancies as concrete audit targets before translating an inventory row into an implementation plan or a claim of superiority. e-model-tablee-datasetse-evaluatione-inconsistenciese-evidence-boundary
5.4 Training and inference
During training
Section 6.1 describes imitation from image–text–action sequences, using an l2 loss for continuous controls or cross-entropy for tokenized actions. Its broad workflow is generic backbone pretraining, multimodal alignment, targeted corner-case/instruction augmentation, and deployment adaptation or compression. This is a synthesis of practices, not a fixed schedule shared by every listed model. e-training
The survey groups PPO, DPO and GRPO under reinforcement fine-tuning, and mentions LoRA, sparse experts and distillation in its deployment stage. It does not define these objectives, their weights, a universal frozen-module policy or a compute budget. Its observation that joint text/action optimization remains rare limits any inference that plausible explanations were directly optimized for safe control. e-training
During inference
A modular language planner passes intermediate outputs to an action head or controller. Unified and augmented systems generate actions or trajectories, sometimes alongside rationales. Figure 1’s Database-to-VLM arrow is explicitly a fine-tuning relation, not evidence of an online retrieval operation in that particular diagram. e-gape-architecturee-stages
The authors separate open-loop trajectory error from closed-loop route success and infractions. Reader analysis: predicting a logged trajectory differs from acting in an environment that responds. Better language answers, imagined futures or offline trajectories cannot by themselves establish safer executed driving. e-evaluation
5.5 Implementation flow
- Ground observations and instructions
Section 3 treats cameras as the primary visual modality, optionally supplemented by other sensors and proprioception. Vision backbones produce latent features, sometimes organized as bird’s-eye-view maps or 3D representations. Language contributes user intent or domain context. Figure 2 abstracts this interface without specifying tensor shapes or mandatory sensor combinations. e-architecture
- Locate language in the decision path
Figure 3 asks whether language merely explains a scene, becomes an intermediate planning representation, participates in a unified policy, or supports reasoning and tool use. Modular action heads consume intermediate representations. Unified systems share perception/planning or caption/trajectory representations. These stages are a survey taxonomy, not controlled architectural ablations. e-stages
- Distinguish decoding from execution
Section 3 lists sequential action/waypoint tokens and diffusion-based continuous outputs among decoder choices. A trajectory may subsequently be executed by an MPC controller. Reader analysis: a diagram terminating in an Action block does not establish the complete feedback loop, control frequency or real-world driving success. e-architecture
- Separate future prediction from reasoning
The survey describes ADriver-I as predicting camera futures conditioned on actions, whereas its ORION description emphasizes memory-based reasoning and candidate trajectories. Reader analysis: these uses of temporal information differ; a VLA label or future-prediction icon alone cannot establish inference-time use of an action-conditioned world model. e-worldmodelse-architecture
6. Experiments & results
This survey organizes autonomous-driving VLA research around how language enters the path from sensing to vehicle action. It distinguishes explanatory overlays, modular language-to-plan interfaces, unified policies and reasoning augmentation. Its contribution is a vocabulary for comparing inputs, representations, action outputs, training and evaluation. It supplies conceptual diagrams and literature inventories rather than a new trained controller or controlled benchmark ranking. Several descriptions conflict across prose, tables and references, so individual model claims require cautious attribution. The architectural and evaluation synthesis is supported by e-gap, e-architecture, e-stages and e-evaluation.
The supplied survey has three conceptual figures and two literature inventories, all included here. Table 2 contains numerical dataset scales, but neither table reports controlled performance comparisons. The source has no original empirical ablation, diagnostic experiment or performance plot, so there is no ablation visual or featured result. Its isolated secondary performance statements lack a complete comparison protocol. The two tables are placed in the results section as survey inventories only; no experiment is inferred from their presence. e-model-tablee-datasetse-evidence-boundary
6.1 Read the original evidence
Table 1. The model inventory helps identify interfaces and provenance, not a winning method. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row across from inputs and data to the two component columns, then output and focus. The source glossary defines Single as a forward-facing camera, Multi as multi-view cameras and State as vehicle state or other sensor information. LLC means low-level control, Traj. means future trajectory and Multi. in the Output column means multiple tasks, which differs from Multi in the Input column. For example, EMMA is listed with multiple tasks, CoVLA-Agent with trajectories and AutoVLA with both low-level control and trajectories. Keep dataset names attached to their rows. Neither a newer year nor a more elaborate focus label is a comparable performance result. e-model-tablee-inconsistenciese-evidence-boundary
What it supports. The table’s fifteen rows make heterogeneity explicit: different sensors, corpora, components and output interfaces are grouped under the same VLA4AD umbrella. That is useful for locating a candidate method and its claimed role. It provides no shared evaluation split, success definition, score, uncertainty or compute budget for ranking those methods.
Where the evidence stops. SimLingo is listed with InternVL2/Qwen-2 here, described as LLaVA-based in Section 4.3, and linked to a bibliography entry titled Carllava. These unresolved source differences prevent treating the row as a verified reproduction specification.
Table 2. Match the evaluation question to the data’s task and domain before comparing its size. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with Domain and Tasks, then inspect Scale. Real denotes recorded real-world data; Sim here denotes CARLA simulation. QA means question–answer pairs and CoT means chain-of-thought. BDD100K/BDD-X is associated with captioning and QA, nuScenes with detection and QA, and Bench2Drive with closed-loop control. Other rows cover graph, multi-turn or action-oriented language tasks. Scale is not one common unit: the table mixes videos, clips, scenes, routes, scenarios, graphs and QA pairs. Section 5 elaborates the benchmark roles, while Section 6.2 distinguishes logged prediction from interactive driving. Use that distinction before choosing what a reported number could actually validate. e-datasetse-evaluatione-inconsistencies
What it supports. Bench2Drive is listed as 220 routes across 44 scenarios, whereas BDD100K/BDD-X is listed as 100 k videos and 7 k clips. Those counts describe different evaluation and annotation resources. The survey supports combining complementary evidence about language and control, but it does not establish that larger tabulated corpora produce safer policies.
Where the evidence stops. DriveLM’s row lists 18 k scene graphs for Real+Sim, while Section 5 separately gives 18 k nuScenes graphs and 16 k CARLA scenes. The table compresses that scope; no reconciled total or split is supplied here.
6.2 Results and evaluation conditions
No quantitative results are included in this reading.
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
No new controlled results or ablations are reported. Secondary claims include TS-VLM BLEU-4 56 and a DiffVLA error-halving statement, but the survey lacks sufficient splits, baseline values, metric details or uncertainty for a reproducible comparison. They are not promoted to structured results. Dataset counts describe inventory scale, not driving performance. e-evidence-boundarye-datasets
Descriptions conflict: DriveGPT-4 is an explainer example but has LLC output in Table 1; SimLingo’s prose and table disagree on its backbone, and reference [71] is titled Carllava. Impromptu-VLA’s table trajectory output differs from the steering/throttle-token prose. DriveLM’s table omits the separate CARLA count given in Section 5. These differences remain unresolved. e-inconsistencies
The authors identify sensor corruption, ambiguous language, hallucinations, regional shift, expensive tri-modal labels and temporal alignment as open problems. Their at-least-30-Hz discussion is a deployment challenge, not a measured speed result. Logic-based vetoes are an initial step toward verification, not proof that general language-conditioned driving is safe. e-challenges
Foundation-scale models, symbolic safety checks and constrained traffic communication are proposed research directions, not evaluated capabilities. Section 8 announces four priority tracks but names only three; no fourth direction is inferred. e-future
7.2 Questions for discussion
- What intervention would show that an explanation causally influences a driving action?
- Which common protocol could score instruction fulfillment and safety without hiding tradeoffs in one aggregate?
- How should a survey record model classifications when its prose and table disagree?
8. Reproducibility audit
8.1 Requirements and known gaps
Individual-model reproduction would require exact sensors, action representation, controller, checkpoint, data/split, objective and simulator configuration. Table 1 supplies only some identifiers; Section 6 provides general practices without optimizer schedules, hardware specifications or complete evaluation settings. The survey alone is insufficient for implementation-level reproduction. e-model-tablee-traininge-evaluation
Reader-proposed checks should first audit table-to-text consistency, then test whether language changes behavior under a fixed driving protocol. Pair identical simulator scenes with equivalent instructions and different valid maneuver requests, keeping the policy and controller fixed while separately scoring instruction fulfillment, route success and infractions. This is a proposal, not a reported experiment. e-inconsistenciese-evaluation
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: Audit the inventory against its own text and bibliography
Reader-proposed consistency check, not an experiment performed here: create a record for every Table 1 row with input, backbone, output type and cited reference number, then compare those fields with Sections 4.1–4.4 and the corresponding bibliography entry. Preserve both values whenever they disagree. Use the DriveGPT-4 output, SimLingo backbone/reference and Impromptu-VLA output as predefined test cases. Separately compare Table 2’s DriveLM scope with Section 5. The falsifiable expectation is internal agreement for each populated field. The already visible differences refute treating these inventories as self-sufficient reproduction specifications; resolution would require the individual primary sources in a separately authorized study. e-model-tablee-inconsistenciese-datasets
Check 2: Test instruction sensitivity without changing the driving setup
Reader-proposed controlled study, not a reported reproduction: select fixed simulator scenes with two feasible, legal maneuver choices. For one documented policy and controller, replay identical initial states and seeds using an instruction, a meaning-preserving paraphrase, a different valid maneuver request and a neutral-command control. Hold sensing, inference budget and routes fixed. Score command fulfillment, route success, collisions and other infractions separately, and record latency on specified hardware. Appropriate grounding predicts stable behavior under paraphrases and appropriate changes under different requests. If only the generated explanation changes, the action-gap claim remains unaddressed; if compliance rises with infractions, the safety tradeoff remains visible. e-gape-architecturee-evaluatione-challenges
8.3 Reading coverage
Visual audit: Inspected the title/author/affiliation block and CVF version watermark on PDF page 1; paradigm, architecture and progression figures on pages 2, 4 and 5; both complete table layouts and legends on page 6; supporting method, training, dataset, evaluation and limitation text on pages 3–8; and reference [71] on page 11. Cross-checked diagram arrow directions against captions and Sections 2–4, including the database fine-tuning arrow and the modular intermediate representation. Viewed every final crop; widened Table 2 to retain the complete closed-loop-control label. All six text chunks, including the remaining reference pages, were read. No empirical ablation page or controlled-results table exists in the supplied survey; no separate supplement was supplied.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 11. 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, affiliations, version watermark and Abstract
- 1. Introduction
- 2. Development of Autonomous Driving (2.1–2.4)
- 3. Architecture Paradigm of VLA4AD (3.1–3.3)
- 4. Progress of VLA4AD Paradigm (4.1–4.4)
- 5. Datasets and Benchmarks
- 6. Training and Evaluation Strategies (6.1–6.2)
- 7. Open Challenges
- 8. Future Directions
- 9. Conclusion
- References [1]–[117], PDF pp. 9–13
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.
- All six supplied text chunks were read individually. All three figures and both tables were subsequently inspected in the PDF, resolving the extraction-only visual gap.
- The title and catalog author string match the title page. The supplied BibTeX orders those authors differently; metadata follows the title page. This is the CVF open-access accepted workshop version. Its watermark states equivalence to the accepted version apart from the watermark; the separate IEEE proceedings and other revisions were not compared. No numbered revision is established.
- The complete supplied PDF contains no appendix. References were read as bibliography entries; cited papers, code, datasets and evaluation servers were not independently inspected. No experiments were reproduced.
- PDF pages 1–8 and 11 were visually inspected. Remaining bibliography pages were read in text; no separate supplement was supplied.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1 (proceedings p. 4583), title, author/affiliation block and CVF watermark
The title and twenty authors in title-page order match the catalog author string. The watermark identifies an ICCV Workshop open-access version, identical to the accepted version except for the watermark; it points to separate final proceedings. No numbered revision is printed here.
Go to primary source ↓e-gapPDF pp. 1–3 (4583–4585), Abstract, Sections 1–2.4 and Figure 1
The survey motivates an action gap between language-based understanding and vehicle control. Figure 1 contrasts direct end-to-end control, a VLM reasoning/multitask branch, and a multimodal encoder–LLM/VLM–action decoder pathway. The database arrow points into the VLM and is labeled Finetune.
Go to primary source ↓e-architecturePDF p. 4 (4586), Figure 2 and Sections 3.1–3.3
The synthesis covers cameras and other sensors, vehicle state, language commands, latent visual representations, pretrained language processors, autoregressive or diffusion action decoding, and controls or trajectories. Trajectories can be executed by MPC controllers. Figure 2 is a conceptual vision-to-language-to-action diagram, not a specific trained network.
Go to primary source ↓e-stagesPDF pp. 5–7 (4587–4589), Figure 3 and Sections 4.1–4.4
Four stages distinguish explanatory language, modular intermediate representations plus action heads, unified prediction, and reasoning/tool augmentation. Section 4.2 identifies latency and interface failures; Section 4.3 describes shared caption/trajectory or perception/planning representations; Section 4.4 discusses memory and interleaved thought/action decoding.
Go to primary source ↓e-worldmodelsPDF p. 6 (4588), Sections 4.3–4.4 below Table 2
The survey describes ADriver-I as predicting future camera frames conditioned on actions using diffusion, DiffVLA as producing trajectories with sparse waypoint and dense occupancy predictions, and ORION as compressing images and past actions with QT-Former for reasoning and candidate trajectories. These are summaries of cited works.
Go to primary source ↓e-model-tablePDF p. 6 (4588), Table 1, headers, caption glossary and all 15 rows
Table 1 records models from 2023–2025, input modalities, datasets, vision and language components, outputs and focus. LLC denotes low-level control, Traj. future trajectory and Multi. multiple tasks. There is no performance metric column, common split or uncertainty estimate.
Go to primary source ↓e-datasetsPDF p. 6 (4588), Table 2; PDF p. 7 (4589), Section 5
Eight dataset/benchmark rows include Bench2Drive with 220 routes and 44 CARLA scenarios, BDD100K/BDD-X with 100 k videos and 7 k clips, and nuScenes with 1 k scenes of 20 s and six cameras. The text adds a 2M-frame Bench2Drive training set. Units and tasks differ across rows.
Go to primary source ↓e-trainingPDF pp. 7–8 (4589–4590), Section 6.1
The survey describes expert imitation with continuous-control l2 loss or token cross-entropy, targeted augmentation or reinforcement fine-tuning, and deployment adaptation/compression. It lists PPO, DPO and GRPO together, and generic pretraining/alignment/augmentation/compression stages. Joint text/control optimization is described as rare; no single executable training recipe is provided.
Go to primary source ↓e-evaluationPDF p. 8 (4590), Section 6.2
Evaluation axes include closed-loop route success and infractions; open-loop trajectory error, collision rate and instruction-conditioned goals; and sensor, language and distribution stress. The authors say control-centric suites and reasoning tests remain disjoint. The DiffVLA error-halving statement lacks a numerical results table or fully specified protocol.
Go to primary source ↓e-challengesPDF p. 8 (4590), Section 7
Challenges include sensor corruption, language ambiguity and hallucinations, regional shift, scarce image/action/language annotations, camera-centric fusion and temporal consistency. Running a vision transformer and LLM at at least 30 Hz is presented as a challenge, not a demonstrated timing measurement.
Go to primary source ↓e-futurePDF p. 8 (4590), Sections 8–9
Proposed directions include foundation-scale multimodal driving models, neuro-symbolic safety kernels, and communication/social intelligence through constrained traffic intents. Section 8 announces four priority tracks but provides only three named tracks.
Go to primary source ↓e-inconsistenciesPDF p. 5 (4587), Figure 3 and Sections 4.1, 4.3; PDF p. 6 (4588), Tables 1–2; PDF p. 7 (4589), Sections 4.4–5; PDF p. 11 (4593), reference [71]
DriveGPT-4 is an explainer example in Section 4.1 but has LLC output in Table 1. SimLingo is described as LLaVA-based in Section 4.3, listed with InternVL2/Qwen-2 in Table 1, and cites [71], titled Carllava in the bibliography. Impromptu-VLA has trajectory output in Table 1 but steering/throttle tokens in Section 4.4. DriveLM has 18 k graphs in Table 2, whereas Section 5 separately gives 18 k nuScenes graphs and 16 k CARLA scenes.
Go to primary source ↓e-evidence-boundaryPDF pp. 5–8 (4587–4590), Figure 3, Tables 1–2 and Sections 4–9
The work synthesizes architectures, datasets, training and evaluation rather than introducing a model with controlled experiments. Its tables are inventories. Section 5 mentions TS-VLM BLEU-4 56 without a specified evaluation split or comparable baseline; Section 6.2 contains another underspecified relative improvement claim. No ablation, controlled performance table or displayed mathematical model is supplied.
Go to primary source ↓8.5 Primary sources
A Survey on Vision-Language-Action Models for Autonomous Driving ↗
PDF · 9,621 extracted words
Source fingerprint
9080e4a5b0dfbb1fef2cbdeba7c9663b152dced54841ec4210c778dbb34077cc