DynVLA: Learning World Dynamics for Action Reasoning in Autonomous Driving
1. Paper overview
In one sentence: Predicting compact ego and environment dynamics before trajectory tokens improves reported driving benchmarks, but forecasting mistakes can become planning mistakes. e03e04e05e10e14e16e20e22
| At a glance | What to know |
|---|---|
| Research problem | Reader analysis Driving requires anticipating nearby agents and road geometry. The authors argue that linguistic descriptions underspecify these relationships, while dense future images spend computation on background and texture. Dynamics CoT aims to preserve motion relevant to planning with fewer generated tokens. It is a learned forecasting representation, not an explicit physics solver. e02e03e04 |
| Core mechanism | Source description Predicted future transitions explicitly condition action generation, rather than serving only as auxiliary training targets. e05 |
| A key reported result | NAVSIM trajectory planning: 91.7 PDMS, higher better. Table 1 standard benchmark protocol; exact split identifier unreported. AdaThinkDrive and AutoDrive-R2: 90.3 each; human reference: 94.8. A 1.4-point lead over the strongest listed learned baselines, below the human reference. Cross-method comparisons do not isolate Dynamics CoT. e09e10 |
| Reading caution |
Core contributions
- Source description
Predicted future transitions explicitly condition action generation, rather than serving only as auxiliary training targets. e05
Figure 2. Learning a compact transition vocabulary, then using it as the action prefix. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read panel (a) from the bottom upward. Adjacent images enter the Dynamics Encoder with two query groups. Green and yellow branches pass through separate VQ codebooks, while the green ego branch also feeds an action-regression loss. Their combined representation reaches the Dynamics Decoder, conditioned on the current image and BEV. Panel (b) shows a different learning stage: snowflakes mark frozen target tokenizers, and the blue DynVLA backbone learns orange dynamics tokens followed by green action tokens. The upper future-image stack supplies training targets. Equations (5)–(9) confirm that inference instead predicts dynamics from the available observation context. e03e04e05e07
What it supports. The representation is constrained by both reconstruction and measured ego motion, then becomes an explicit input to action generation. The evaluated default uses eight tokens per transition and two transitions, giving sixteen dynamics tokens. The one-policy classification comes from the shared DynVLA generation path, not from treating the entire training pipeline as one network.
Where the evidence stops. The single Dynamics Decoder box summarizes two modality-specific decoders defined in Eq. (3). Its reconstructed future images and the auxiliary ego-action head should not be mistaken for required inference-time rendering or the deployed controller.
2. Motivation
2.1 The problem and the proposed response
Driving requires anticipating nearby agents and road geometry. The authors argue that linguistic descriptions underspecify these relationships, while dense future images spend computation on background and texture. Dynamics CoT aims to preserve motion relevant to planning with fewer generated tokens. It is a learned forecasting representation, not an explicit physics solver. e02e03e04
2.2 What this reading follows
DynVLA asks what a driving policy should predict before choosing a trajectory. Its answer is a small discrete description of scene change: one part for ego motion and another for surrounding dynamics. A tokenizer learns this representation from adjacent frames, then a single policy predicts it before generating actions. This edition follows the training-to-inference handoff, separates final benchmark results from controlled ablations, and examines why more prediction is not always better. The source provides strong reported comparisons and useful failure examples, while leaving uncertainty estimates, exact evaluation splits and some implementation settings unresolved. e03e04e05e10e14e16e20e22
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 | One Model |
| Prediction paradigm | IDM |
| Quadrant | Q2 · One Model × IDM |
3.1 Evidence-based assessment
Supports the recorded classification
One Model holds at the inference-policy level: one causal DynVLA backbone generates dynamics and then actions. IDM is supported as future-conditioned action generation, with no separate inverse-dynamics network required at inference. The tokenizer and auxiliary action decoder are training machinery. Latent prediction is supported; a JEPA objective is not described. e03e04e05
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 reconstructing a future can still produce useless tokens
The tokenizer is trained from image pairs, but its decoder also sees the current state. That shortcut is useful: it lets the decoder recover background without forcing a tiny code sequence to store the whole scene. It also creates a failure mode. If reconstruction is dominated by information already available in the current observation, the dynamics codes may carry little useful change. DynVLA addresses this with separate query/codebook branches and an ego-action regression target. The same codes must additionally support both image and BEV reconstruction. Figure 5 and Table 6 connect this bundled design to greater code utilization and improved planning. They support the design as a package, while leaving the independent contribution of query separation versus action supervision unresolved. e03e04e15
5.2 Turn hindsight targets into a foresight-conditioned policy
During target construction, the tokenizer can inspect the actual current and future frames. Those codes summarize what happened; they are not predictions yet. SFT teaches DynVLA to predict the codes from observation history, instruction and ego state, then to predict actions conditioned on that prefix. Equation (8) makes the dependence explicit: the action likelihood sees the complete dynamics sequence. The default uses two one-second transitions, giving sixteen dynamics tokens before action generation. Frozen target tokenizers in Figure 2 should therefore be read as teachers, not as a requirement to observe the future at deployment. This distinction also explains the taxonomy: one policy performs future-conditioned action generation, while separate training modules define its intermediate representation. e03e05e07
5.3 Reward the trajectory while keeping the reasoning format
SFT teaches the order and content of the output, but its action targets remain imitation-based. RFT evaluates sampled sequences with PDMS and a binary format reward, standardizes rewards within a group, and updates the policy using GRPO with clipping and KL to the frozen SFT reference. The trajectory score rewards the outcome; the format reward only checks that the required organization is preserved. Table 5 helps separate these effects empirically: EMU3 improves from 85.6 to 88.7 with RFT without Dynamics CoT, while the dynamics-conditioned model rises from 87.2 to 91.7. This supports complementarity in the reported runs. It does not certify that every generated intermediate forecast is correct, especially given the failures acknowledged in Appendix E. e06e09e14e20
5.4 Training and inference
During training
Tokenizer training combines image MSE/LPIPS, BEV cross-entropy, VQ loss and squared ego-action error. Hidden width is 1024, with a 12-layer encoder, two 8-layer decoders and 32-dimensional VQ embeddings. BEV weight is 0.1; other listed loss weights are 1. e04e07e08
EMU3 is the default backbone. SFT equally weights dynamics/action likelihood losses. GRPO samples candidate sequences, standardizes group rewards and uses clipped policy updates plus KL to a frozen SFT reference. PDMS trajectory and binary format rewards each have weight 1. e05e06e08
Tokenizer/SFT/RFT use 200k/4k/6k steps on 8 L20/8 L20/6 H800 GPUs. Peak learning rates are 1e-4/1e-4/2e-6. These are reported configurations, not measured reproduction costs. e08
During inference
Generate future dynamics, then action tokens, from the observation/instruction/state context. Two one-second transitions contain sixteen dynamics tokens before delimiters. Figure 1’s eight-token shorthand describes the per-transition budget, not the complete default prefix. e02e05e07
Future images supply training targets; inference predicts their dynamics from available context. Image/BEV reconstruction and the ego regularization head are training/visualization machinery. Dense future-image generation is not required by the described action path; controller details and replanning cadence remain unspecified. e04e05e22
5.5 Implementation flow
- Encode transitions
Patchify adjacent images and aggregate change with ego and environment query sets. Quantize each branch with a separate nearest-neighbor codebook, then concatenate the codes. Defaults are four tokens per branch and 64 entries per codebook. e03e07
- Constrain the bottleneck
Decode shared dynamics embeddings into future image and BEV, conditioned on their current representations. A two-layer MLP predicts relative ego motion from ego codes. Current-state conditioning reduces static-information demands; action supervision discourages uninformative dynamics. e04e15
- Learn the ordered sequence
Frozen tokenizers supervise dynamics followed by FAST actions, with start/end markers for each span. The causal policy learns both spans; action predictions see the dynamics prefix. e05
6. Experiments & results
DynVLA represents driving foresight with discrete ego-motion and environment-motion tokens, then conditions trajectory generation on them. A separately trained tokenizer supplies targets; one autoregressive policy learns dynamics-first action generation and undergoes reinforcement fine-tuning. Reported NAVSIM PDMS reaches 91.7. The tradeoff is compact prediction versus propagation of prediction errors into planning, with benchmark gains but incomplete reproduction details and no demonstrated physical-road deployment (e03–e16, e20–e22).
6.1 Read the original evidence
Table 1. Final-model NAVSIM scores, with the human reference retained. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the shaded PDMS column and compare DynVLA with the best listed learned methods, then return to the component columns. NC is no-at-fault collision, DAC is drivable-area compliance, TTC is time-to-collision, C is comfort, and EP is ego progress. Appendix A combines these measures into trajectory-level PDMS; the table displays aggregate results. The grouping distinguishes model families, not matched training budgets. Read the gray human row separately from learned baselines. Bold and underline identify the paper’s best and second-best results; they do not indicate confidence intervals or statistical significance. e09e10e14
What it supports. DynVLA reports 91.7 PDMS, 1.4 points above the 90.3 entries for AdaThinkDrive and AutoDrive-R2, while the human reference remains higher at 94.8. Its 86.8 progress score accompanies strong safety/compliance scores, but it does not lead every component column. This is the final-model result, distinct from the 87.2 SFT ablation.
Where the evidence stops. The source does not name the exact evaluated split here or establish equal pretraining and compute across methods. Benchmark PDMS is not a physical-road collision rate, and these point estimates provide no run-to-run uncertainty.
Figure 10. When an incorrect imagined transition guides the next action. Original paper, p. 18 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each visible row from current observation to decoded future and ground-truth future, then inspect the planning map at the left. The upper example concerns a leading vehicle; the middle example shows a turning scene; the bottom example has degraded rainy imagery. The point is the disagreement between the imagined and observed future, not the visual sharpness of the reconstruction. Figure 6 provides the shared map key, including human and planned trajectories. Appendix F discusses agent-intention errors, drivable-area errors and degraded observations, but its row numbering does not align with the three-row graphic supplied here. e19e20e21
What it supports. These selected examples illustrate the vulnerability of an explicit dynamics prefix: a mistaken forecast can accompany an unsuitable trajectory. This supports the authors’ stated concern about error propagation and motivates testing the dependence of actions on prefix content. The examples establish possible failure modes, not their prevalence in the benchmark.
Where the evidence stops. Appendix F refers to four rows, while Figure 10 has three; the exact row mapping remains unresolved. No standalone legend or failure frequency is supplied in Figure 10. The crop preserves the original layout rather than inventing a missing example.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| NAVSIM trajectory planning Table 1 standard benchmark protocol; exact split identifier unreported. | 91.7 PDMS, higher better | AdaThinkDrive and AutoDrive-R2: 90.3 each; human reference: 94.8. A 1.4-point lead over the strongest listed learned baselines, below the human reference. Cross-method comparisons do not isolate Dynamics CoT. e09e10 |
| Bench2Drive closed-loop driving Interactive benchmark; exact evaluated route set/run counts unreported. | 88.34 / 72.73 / 72.23 Driving Score / Success Rate / mean Multi-Ability | SimLingo: 85.07 / 67.27 / unreported. Privileged PDM-Lite: 97.02 / 92.27 / 92.82. Best among reported nonprivileged entries, not all rows. Simulated closed-loop performance does not establish physical-road safety. e09e11 |
| In-house three-second trajectory prediction Training on 700k in-house frames; held-out split size/access unspecified. | 1.215 / 4.04 ADE (m) / collision rate (per ten thousand) | DriveVLA-W0 (ViT): 1.344 / 5.13. Lower displacement error and predicted-trajectory collision rate, not measured deployed collision statistics. e09e12 |
| CoT quality and inference latency Table 4, single NVIDIA H800; dynamics row matches Table 5 SFT-only performance. | 87.2 / 0.37 PDMS / latency (s) | No CoT: 85.6 / 0.20; scene description: 85.3 / 3.04; future image: 86.3 / 2.29. Calculated total-latency speedups are approximately 8.2× and 6.2× over descriptions and images, below the introduction’s >10× claim. Dynamics adds 0.17 s over direct prediction. e13e14 |
| Dynamics CoT and reinforcement fine-tuning Table 5, controlled stage comparisons on two backbones. | EMU3 SFT without/with Dynamics CoT: 85.6/87.2; SFT+RFT without/with Dynamics CoT: 88.7/91.7. PDMS | Qwen2.5-VL corresponding scores: 85.3/86.6 and 88.4/91.0. Dynamics conditioning and reinforcement refinement are complementary in these runs; uncertainty across seeds is absent. e14 |
| Tokenizer design and prediction horizon Tables 6–8, planning ablations without reported run-to-run uncertainty. | Full tokenizer: 87.2; without decoupling: 85.8. K=2: 87.2 at 0.37 s; K=4: 86.6 at 0.61 s. PDMS and latency | Decoupled BEV-only: 86.2; image-only: 86.7. Balanced eight-token allocation beats tested four/sixteen-token alternatives. More capacity/foresight does not monotonically help. Decoupling changes queries and action supervision together, leaving individual effects unresolved. e15e16e17 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 4. Compact dynamics improves the reported quality/latency tradeoff. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the latency and PDMS columns together rather than ranking either in isolation. The first row provides the no-reasoning baseline. Scene Description and Meta Action encode different textual abstractions; Future Image and Optical Flow use denser visual outputs. Dynamics supplies the compact learned transition representation. The caption specifies a single NVIDIA H800 GPU for latency measurement. Its 87.2 PDMS matches the Dynamics CoT SFT row of Table 5, so this comparison should not be combined with the final 91.7 score as though both described one measured operating point. e13e14
What it supports. Dynamics improves PDMS from 85.6 to 87.2 at an added 0.17 seconds over no CoT. Its reported latency is approximately 8.2 times lower than scene descriptions and 6.2 times lower than future images, calculated from the displayed values. It remains slower than direct action prediction despite being faster than the compared reasoning variants.
Where the evidence stops. The introduction claims more than an order-of-magnitude latency reduction, which these total latencies do not establish. The table does not give uncertainty, workload variation or a complete latency breakdown, so the ratios describe this reported measurement only.
Table 6. The gain depends on how dynamics is learned, not merely on adding tokens. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the three design columns as a checklist. The first Dynamics CoT row retains image and BEV reconstruction but omits Decouple. The next row uses decoupling with BEV only; the following row uses decoupling with image only; the final row enables all three. Section 4.3 defines Decouple as both separate ego/environment queries and ego-action regularization. Compare these rows against the no-CoT reference above them. Figure 5 on the same page adds a diagnostic: activated-code counts remain much lower without the bundled decoupling design, consistent with its weak planning improvement. e04e15
What it supports. With both reconstruction branches but no decoupling, PDMS is 85.8, close to the 85.6 no-CoT baseline. The complete tokenizer reaches 87.2. BEV-only and image-only variants with decoupling reach 86.2 and 86.7 respectively. Thus the reported best setting combines the physical supervision/factorization bundle with both reconstruction views.
Where the evidence stops. The Decouple switch changes two mechanisms together; this table cannot assign the gain separately to queries or action supervision. Activated-code counts diagnose utilization, but neither code utilization nor these planning scores alone proves semantic disentanglement.
Table 7. Two seconds is the best tested lookahead in this ablation. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The leftmost column varies K, the number of predicted future transitions, not the length of the action-token sequence. Appendix C specifies one-second spacing, so K equals two covers two seconds. At eight dynamics tokens per transition, the default produces sixteen dynamics tokens; extending K also increases the generated prefix. Compare each row’s latency with its PDMS to see the cost of longer reasoning. The no-CoT row anchors the comparison. Table 8, also on this page, studies a different axis—the per-transition token budget and its ego/environment allocation—which should not be confused with K. e07e16e17
What it supports. PDMS rises from 86.5 at K=1 to 87.2 at K=2, then falls to 86.7 and 86.6 for K=3 and K=4. Latency increases from 0.27 to 0.61 seconds across the tested horizons. Longer prediction therefore fails to buy better average planning in this configuration, motivating the paper’s two-second default.
Where the evidence stops. The authors attribute the decline to uncertain long-range dynamics, but the table does not directly measure uncertainty. Horizon and generated-token count change together, and the absence of repeated-run statistics limits interpretation of the smaller score differences.
7. Analysis & limitations
7.1 What the evidence leaves open
Inaccurate agent intentions, novel-view geometry and degraded observations can corrupt dynamics and subsequent actions. Appendix F’s four-row references conflict with the three visible rows in Figure 10; exact row mapping remains unresolved. e20e21
Selected transfer/reconstruction examples do not quantify disentanglement or establish that faithful dynamics cause planning gains. The coupled decoupling ablation confounds architecture and supervision. e15e18e19
Exact splits, route/sample counts, software versions and statistical uncertainty are missing. The fast/slow controller is proposed future work; the experiments do not demonstrate on-road deployment. e09e20e22
7.2 Questions for discussion
- Does action generation depend on correct dynamics, or would an equally long uninformative prefix retain the gain? (e05, e14)
- How much tokenizer improvement comes from separate queries versus ego-action supervision? (e15)
- Can prediction uncertainty identify failures before unsafe actions are selected? (e20, e21)
8. Reproducibility audit
8.1 Requirements and known gaps
Require the referenced EMU3 pretraining, FAST, temporal front images, ego-motion labels and BEV supervision. Match two one-second transitions, four tokens per branch and 2048 action-token types. AdamW betas are 0.9/0.95, batches 32/6/6, cosine warm-ups 1000/100/500 steps; RFT accumulation is 6 and KL coefficient 1e-3. e07e08
A faithful rerun needs exact splits, the external pretraining recipe, GRPO group size/clipping, sampling settings and controller details. Proposed checks should separate query/action-supervision factors and intervene on the dynamics prefix; neither has been run. e05e06e15e22
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: Separate query factorization from action supervision
Reader-proposed check, not performed: run a two-by-two tokenizer experiment with shared versus branch-specific queries and ego-action regularization off versus on. Keep total query count, total codebook capacity, both reconstruction branches, data, optimizer and downstream SFT budget fixed; leave RFT off initially. Across matched seeds, compare activated-code counts, held-out ego-motion prediction error and planning PDMS. If action supervision alone recovers the utilization and planning gain, separate queries are not independently necessary under this setting. If only the combination works, the interaction is the meaningful result. This resolves a confound in Table 6 rather than repeating its bundled switch. e03e04e08e15
Check 2: Intervene on the predicted dynamics prefix
Reader-proposed check, not performed: freeze the same K=2 SFT policy and compare its normal sixteen-token prefix with shuffled ego tokens, shuffled environment tokens, and oracle tokens extracted from the actual future. Preserve prefix length, delimiters and action decoding settings; use the same held-out scenes and report PDMS components plus intention/turning/degraded-view subsets. Oracle futures are a diagnostic upper bound, not deployable input. If branch-specific corruption leaves actions and scores unchanged, the claimed role of meaningful dynamics is weakened. If oracle prefixes improve scores and targeted corruption produces corresponding failures, that supports useful content dependence, while still not proving complete disentanglement. e05e07e09e14e20e21
8.3 Reading coverage
Visual audit: Inspected the title/author page; all ten figures; all eight tables; method, objective, protocol and appendix implementation pages supporting retained claims. Every final original crop was inspected at its actual rendered size. Equations and training-stage diagrams were cross-checked for generation order and frozen modules. Figure 2’s summarized decoder was reconciled with Eq. (3); Figure 1’s token shorthand was distinguished from Appendix B’s default sequence. Appendix F/Figure 10 row mismatch is disclosed. Pages 9–12 contain references and were read in full as text. No separate supplement, external code or experiment execution was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 13, 14, 15, 16, 17, 18. 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
- 1. Introduction
- 2. Related Works (2.1–2.2)
- 3. Method (3.1–3.3)
- 4. Experiments (4.1–4.4)
- 5. Conclusion
- Impact Statement
- References
- Appendix A. Datasets and Metrics
- Appendix B. Implementation Details
- Appendix C. More Ablation Studies
- Appendix D. More Qualitative Comparisons
- Appendix E. Limitations and Future Works
- Appendix F. Failure Cases
- Appendix G. More Related Works (G.1–G.2)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This acquisition limitation was addressed by inspecting every figure and table in the supplied PDF.
- Separate supplemental material availability has not been fully verified.
- Identity/version: the exact title and all twelve catalog authors match the inspected title page. This is arXiv:2603.11041v2, stamped 13 March 2026, with a preprint footer dated 16 March 2026. The catalog’s ICML 2026 venue is not established by this preprint; no venue edition or earlier revision was supplied for comparison.
- All seven text chunks were read, covering all eighteen pages. Visual inspection covered pages 1–8 and 13–18; reference-only pages 9–12 were read as text.
- Code, external project material and separate supplements were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title, author/affiliation block, arXiv margin and preprint footer
Exact title and all twelve authors match. The margin says arXiv:2603.11041v2, 13 Mar 2026; the footer says Preprint, March 16, 2026. Affiliations are CASIA and Yinwang Intelligent Technology Co. Ltd.
Go to primary source ↓e02PDF pp. 1–2, Abstract, Figure 1 and Section 1
Dynamics CoT is proposed as compact future-dynamics reasoning between observations and actions, contrasting textual descriptions and dense image prediction. Figure 1 labels eight dynamics tokens.
Go to primary source ↓e03PDF p. 3, Figure 2(a), Section 3.1, Eqs. (1)–(2)
Adjacent images and separate ego/environment queries feed a Transformer encoder. Separate nearest-neighbor VQ codebooks produce concatenated dynamics tokens.
Go to primary source ↓e04PDF p. 4, Section 3.1, Eqs. (3)–(4) and action-regularization paragraph
Shared dynamics embeddings reconstruct future image/BEV conditioned on the current representations. A two-layer action decoder supervises ego codes with squared relative-motion error. The objective combines reconstruction, VQ and action regularization.
Go to primary source ↓e05PDF pp. 3–5, Figures 2(b) and 3, Section 3.2, Eqs. (5)–(9)
Frozen tokenizers supervise dynamics before FAST action tokens. Context is current/previous images, text and ego state. Both spans have start/end markers; action likelihood conditions on the full dynamics prefix.
Go to primary source ↓e06PDF pp. 4–5, Figure 3 Stage 3, Section 3.3 and Eq. (10)
GRPO samples sequences, uses group-standardized trajectory/format rewards, clipped policy ratios and KL to a frozen SFT reference. Trajectory PDMS is in [0,1].
Go to primary source ↓e07PDF pp. 13–14, Appendix B, Dynamics Tokenizer and Dynamics CoT SFT
Each transition has four ego/four environment tokens and two 64-entry codebooks. VQ dimension is 32, hidden width 1024, encoder 12 layers, decoders 8 layers. Two one-second transitions yield sixteen tokens. FAST has 2048 types; inputs use current and one-second-old front views.
Go to primary source ↓e08PDF p. 14, Appendix B, all training-stage paragraphs
Tokenizer/SFT/RFT: 200k/4k/6k steps; 8 L20/8 L20/6 H800 GPUs; peak rates 1e-4/1e-4/2e-6; warm-ups 1000/100/500; batches 32/6/6; AdamW betas 0.9/0.95. RFT accumulation is 6, KL 1e-3. BEV weight is 0.1, other listed weights 1.
Go to primary source ↓e09PDF p. 13, Appendix A, all dataset paragraphs and Eq. (11)
PDMS combines NC, DAC, EP, TTC and comfort. Bench2Drive measures destination success, penalized driving score and five-ability mean. The in-house dataset has 700k frames and a three-second displacement/collision evaluation; exact splits/counts are not specified.
Go to primary source ↓e10PDF p. 6, Table 1, DynVLA, Human, AdaThinkDrive, AutoDrive-R2 and component columns
DynVLA PDMS is 91.7; strongest listed learned comparators are 90.3, Human 94.8. DynVLA components are NC 98.6, DAC 98.7, TTC 95.5, C 100 and EP 86.8.
Go to primary source ↓e11PDF p. 6, Table 2, DynVLA, SimLingo, TF++, privileged rows and dagger definition
DynVLA reports 88.34 DS, 72.73 SR, 72.23 mean ability. SimLingo has 85.07/67.27, no ability mean; TF++ ability mean is 64.39. Privileged Think2Drive and PDM-Lite exceed DynVLA; PDM-Lite has 97.02/92.27/92.82.
Go to primary source ↓e12PDF p. 6, Table 3 and in-house paragraph; p. 13, Appendix A, In-House Dataset
DynVLA ADE/collision: 1.215 m/4.04 per ten thousand; DriveVLA-W0 ViT: 1.344 m/5.13. Comparison methods are reimplemented; collisions concern predicted trajectories within three seconds.
Go to primary source ↓e13PDF p. 6, Table 4 and caption; p. 2, Section 1 latency claim
Single-H800 latency/PDMS: no CoT 0.20/85.6; description 3.04/85.3; meta-action 0.43/86.0; image 2.29/86.3; flow 2.29/86.4; dynamics 0.37/87.2. Total-latency ratios do not reach the introduction’s claimed order-of-magnitude reduction.
Go to primary source ↓e14PDF p. 6, Table 5, checked stage columns for EMU3 and Qwen2.5-VL
EMU3 SFT without/with Dynamics CoT scores 85.6/87.2; adding RFT gives 88.7/91.7. Qwen2.5-VL follows the same ordering: 85.3/86.6 and 88.4/91.0.
Go to primary source ↓e15PDF p. 7, Figure 5, Table 6 and Section 4.3; p. 8, continued ablation analysis
Full tokenizer PDMS 87.2; no decoupling with both reconstruction branches 85.8; decoupled BEV-only 86.2; image-only 86.7. Decoupling jointly introduces separate queries and action regularization. Figure 5 compares activated-code trajectories.
Go to primary source ↓e16PDF p. 14, Appendix C, Table 7 and horizon analysis
K=1/2/3/4 gives latency 0.27/0.37/0.49/0.61 seconds and PDMS 86.5/87.2/86.7/86.6. Transitions are one second apart; K=2 is default.
Go to primary source ↓e17PDF p. 14, Appendix C, Table 8 and token-allocation analysis
Eight tokens with 4/4 allocation score 87.2; 2/6 and 6/2 score 86.9/86.4. Four total tokens score 86.4; sixteen score 86.5.
Go to primary source ↓e18PDF p. 7, Figure 4 and caption; p. 16, Figure 8
Selected cross-scene transfers change ego/agent motion in decoded image/BEV, providing qualitative transferability demonstrations.
Go to primary source ↓e19PDF p. 8, Figure 6 and Section 4.4; p. 15, Figure 7 and Appendix D; p. 17, Figure 9
The figures compare planning with/without Dynamics CoT and tokenizer reconstructions. Figure 6 supplies the semantic/trajectory legend, with green human and red planned trajectories.
Go to primary source ↓e20PDF pp. 15–16, Appendix E
The authors acknowledge that inaccurate future dynamics can induce action errors, proposing richer priors and a future fast/slow architecture.
Go to primary source ↓e21PDF p. 16, Appendix F; p. 18, Figure 10 and caption
Failures involve agent intentions, turning/drivable-area mistakes and degraded rainy observations. The figure has three rows; Appendix F refers to first/second, third and fourth rows, an unresolved inconsistency.
Go to primary source ↓e22PDF pp. 13–14, Appendices A–B; p. 5, Section 3.3
Implementation refers to external pretraining, but does not provide exact splits, software versions, GRPO group size/clipping, full controller details or statistical uncertainty.
Go to primary source ↓8.5 Primary sources
DynVLA: Learning World Dynamics for Action Reasoning in Autonomous Driving ↗
PDF · 10,981 extracted words
Source fingerprint
35129e87321794e7bfdcf4ad34e6658fdda7d9bf8117988443ac8243060dcc64