Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features
1. Paper overview
In one sentence: Intermediate video features can support strong driving plans, while a learned quality verifier trades backbone depth for latency without a safety guarantee. e-interfacee-methode-routinge-auxiliarye-routing-resultse-safety
| At a glance | What to know |
|---|---|
| Research problem | Source description Driving requires an ego trajectory, yet video-based planners can inherit expensive future synthesis or a fixed backbone depth. The paper asks whether planning quality can determine how much of a pretrained video model must run, while retaining predictive supervision during learning. e-interfacee-method |
| Core mechanism | |
| A key reported result | NAVSIM v1 adaptive planning and latency: 90.79 PDMS (90.8 rounded); 170 ms; 94.1% exit by B15. PDMS ↑; mean end-to-end latency ↓. Navtest, front camera without LiDAR; one trajectory per attempted exit; threshold 90; A100 80GB, batch one. Fixed B15: 90.62 / 190 ms; full path: 85.82 / 320 ms. The gain over B15 is 0.17 points with about 10% lower mean latency. The full path additionally decodes once; the 47% reduction is not an isolated depth-only comparison. e-v1e-routing-resultse-latency |
| Reading caution | Reader analysis NAVSIM v1 uses non-reactive background actors; v2 approximates feedback with precomputed follow-up scenes. The authors explicitly deny a safety guarantee from the stopping threshold. e-protocole-v2e-safety |
Core contributions
Figure 4. A video-trained backbone supplies several action exits; the verifier controls how far planning proceeds. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the image and text inputs at the upper left, then follow the blue Wan path across depth. Dashed blue readouts supply trajectory features at blocks 5, 9, 15, 18, 22 and 30; the repeated action heads are compressed into one schematic box. The upper-right future-scene branch represents training supervision, as the caption states. Deployment follows the trajectory and scorer path without the depicted video sampling loop. Interpret the lower routing graphic using Algorithm 1: score each new candidate, retain the best accumulated candidate, and either stop or resume the cached prefix. Appendix D.4 also clarifies that LoRA spans Wan, beyond the pictured readout locations. e-architecturee-methode-routinge-traininge-optim
What it supports. The important separation is between learning predictive visual features and rendering future pixels online. Each attempted exit still runs a five-step trajectory diffuser and a quality scorer. Saving backbone work therefore depends on sufficiently early acceptance; attaching several heads does not by itself make every input cheaper.
Where the evidence stops. The diamond shows >, but Algorithm 1 uses ≥. Its dashed NO path heads toward the output region rather than clearly returning to the next exit. Follow Eq. (6) and Algorithm 1 for continuation and accumulated-best selection; the faithful schematic is not a complete execution graph.
2. Motivation
2.1 The problem and the proposed response
Driving requires an ego trajectory, yet video-based planners can inherit expensive future synthesis or a fixed backbone depth. The paper asks whether planning quality can determine how much of a pretrained video model must run, while retaining predictive supervision during learning. e-interfacee-method
2.2 What this reading follows
A driving policy needs a trajectory, but a video diffusion model is built to spend computation generating future images. Adaptive-WAM separates those requirements. Video prediction remains a training objective, while deployment reads a conditional Wan prefix and tries trajectory heads at selected depths. A small verifier asks whether the best plan seen so far is adequate. The central evidence is therefore the relationship between quality, exit depth and measured planning cost. Read the adaptive result separately from the stronger 64-proposal comparator: they use different decoders and supervision. The figures below connect that distinction to the architecture, benchmark rows and diagnostic evidence. e-interfacee-methode-routinge-auxiliarye-routing-resultse-safety
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 | Joint prediction |
| Quadrant | Q3 · Dual-system × Joint prediction |
3.1 Evidence-based assessment
Insufficient evidence to decide
Distinct Wan and trajectory-diffusion modules, plus a separately optimized verifier, support Dual-system. Joint video/trajectory supervision is explicit during training, but deployed inference predicts trajectories from video features without jointly generating a visual future or performing inverse dynamics. Thus “Joint prediction” and Q3 need a training-versus-inference qualification; the catalog snapshot is preserved. e-interfacee-methode-traininge-routing
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 Separate the video-noise index from action denoising
There are two diffusion processes to keep distinct. The video backbone is evaluated once at fixed video-noise index 17; this is neither a completed video rollout nor seventeen sequential sampling iterations. Its current-image latent slice is conditioned on the observation, while its future portion starts from scheduler noise. Intermediate hidden states then become inputs to independent action diffusers. Every attempted action head performs five DDIM updates and emits eight ego poses. Training supplies future frames to the video objective, but deployment supplies only the current image and observation-derived description to Wan. The description uses current or historical attributes, with cached text embeddings. Thus the method retains a representation learned through prediction while removing the repeated conditional/unconditional video-synthesis loop from the online planning path. e-methode-datae-traininge-latency
5.2 Train the verifier without letting it reshape proposals
The scorer predicts evaluator components from an image and a candidate trajectory. It is optimized using soft-label BCE rather than a strict ranking loss because many candidates receive tied or saturated metric values. Its image encoder is fine-tuned, but generated trajectories are detached: scorer gradients cannot move their coordinates or alter the Wan actor. Actor and scorer updates alternate during imitation learning. Planner-only RL then freezes the scorer and Wan while refining the trajectory heads. Reader interpretation: this separation makes the verifier's role easier to identify, but also makes calibration after actor refinement a useful reproduction target. The reported 91.2% exact selection accuracy concerns a fixed offline candidate pool; it does not directly measure the false-accept rate of threshold 90 on the evolving, at-most-six-candidate routing pool. e-traininge-scorere-optime-routinge-method
5.3 Distinguish the stopping depth from the chosen trajectory
At B5, the planner has one trajectory and its predicted component score. If that score is insufficient, it resumes the cached Wan state to reach B9, produces another trajectory and compares the two. This continues through the selected exits. When the best accumulated predicted score reaches the threshold, or when B30 is reached, the controller returns the accumulated best. Reader interpretation: because previous scores remain cached, a first threshold crossing must come from the new candidate; an earlier trajectory can be returned by the forced final fallback. The policy therefore searches representations as well as deciding computation. Its cost includes every attempted trajectory head and scorer, not merely the last backbone block. The threshold sweep tests the combined policy; the overlap heatmap only motivates why multiple depths might offer useful alternatives. e-routinge-methode-overlape-routing-results
5.4 Training and inference
During training
Filter NAVSIM navtrain into 82,555 complete nine-frame clips from 103,288 candidates. Future frames provide four seconds of video supervision. Wan uses 1280×704 images; the scorer uses 832×480. Descriptions use current or historical attributes and cached text embeddings. e-data
During imitation learning, shared Wan LoRA, projections and action heads learn from video and trajectory losses. Base Wan weights and VAE remain frozen. Alternating scorer updates use detached generated trajectories with equal-weight soft-label BCE, preventing scorer gradients from changing the actor. e-traininge-optim
Planner-only DiffGRPO subsequently freezes Wan and the scorer and updates action heads. Each five-step denoising chain receives evaluator reward; groups contain eight sampled trajectories, with a behavior-cloning term. These training groups are distinct from the one-per-exit inference pool. e-traininge-optime-method
During inference
The deployed pool holds at most six trajectories. Cached-prefix continuation avoids restarting Wan, but every attempted exit still pays for action denoising and scoring. The selected threshold is 90 on a 100-point scale; video classifier-free sampling and future-video decoding are skipped. e-methode-routinge-routing-results
Figure 4 shows “Score > Threshold?”, while Eq. (6) discussion and Algorithm 1 specify greater-than-or-equal. This report follows the explicit algorithm. Its future-scene branch and LoRA placements are schematic; LoRA actually spans the backbone. e-architecturee-routing
5.5 Implementation flow
- Read a conditional video prefix
Wan2.2-TI2V-5B processes the current image and description at index 17 of a 40-step video schedule. The future latent portion starts from scheduler noise, not observed future frames. e-method
- Decode independent action exits
Blocks 5, 9, 15, 18, 22 and 30 feed separate projections and ReCogDrive-style diffusion heads. Each head conditions on video features, ego state and navigation, then uses five DDIM updates to produce one trajectory. e-method
- Verify accumulated quality
A fine-tuned DINOv2-Small image encoder and trajectory MLP feed six component heads: no-at-fault collision, drivable-area compliance, driving-direction compliance, time-to-collision, ego progress and comfort. The scorer receives neither ego state nor route command. e-scorer
- Continue only when necessary
Cache the best predicted trajectory and continue from the previous hidden state when its score is insufficient. Return the accumulated best at the threshold or at B30. A forced final fallback can return a trajectory from an earlier exit. e-routing
6. Experiments & results
Adaptive-WAM turns a video-diffusion backbone into a driving planner that stops when an intermediate trajectory looks adequate. Video prediction shapes training; deployment uses one conditional feature pass and a separate quality verifier. Its adaptive result is 90.79 PDMS at 170 ms, while the higher 92.6 score belongs to a different fixed-exit, 64-proposal system.
6.1 Read the original evidence
Table 2. Two Adaptive-WAM rows represent different planning systems and proposal budgets. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Locate the two Adaptive-WAM rows at the bottom before comparing the bold final column. The adaptive row decodes one trajectory at each attempted exit; the fixed-B22 row generates 64 proposals at one depth using a separate decoder and coverage supervision. C means camera, while CL adds LiDAR. Read NC as no-at-fault collision, DAC as drivable-area compliance, TTC as time-to-collision, Comf. as comfort and EP as ego progress. PDMS combines these components through the multiplicative and weighted rule in Appendix B.1. The component columns matter because a larger aggregate does not require improvement in every constituent metric. e-v1e-protocole-auxiliarye-routing-results
What it supports. The adaptive row reports 90.8 PDMS against the listed DriveVA value of 90.5, with higher progress, 85.1 versus 83.5, but lower NC and TTC. The 92.6 result demonstrates the auxiliary proposal system's performance. It does not establish that adaptive routing itself reaches that score.
Where the evidence stops. Table 2 sources baseline values through the cited comparison; equal training data, camera coverage and proposal budgets are not established here. The auxiliary model changes both decoder and supervision. Its score must not be combined with the adaptive model's 170 ms latency.
Table 5. Threshold 90 gives the strongest reported adaptive score; greater depth expenditure is not uniformly better. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the fixed-B15 row as the principal reference: one intermediate depth already gives 90.62 PDMS at 190 ms. Then follow the adaptive rows as the quality threshold rises. Eta is a predicted score threshold on the 100-point scale, not a probability of safe driving. 'Exit by B15' includes termination at B5, B9 or B15 and does not mean that all those scenes use B15. Each attempted exit contributes one candidate; the returned candidate can come from an earlier depth. The caption and Appendix K define latency as batch-one planning time on one A100 80GB, with cached text and all attempted heads and scorer evaluations included. e-routinge-routing-resultse-latency
What it supports. At eta 90, 94.1% of scenes stop by B15, yielding 90.79 PDMS at 170 ms. Eta 70 is faster at 112 ms but drops to 88.49. Raising eta to 95 costs 284 ms while reaching 90.75. The reported operating points show a quality–latency trade-off, not monotonic improvement with stricter acceptance.
Where the evidence stops. The full path includes one VAE decode, so its 320 ms cost does not isolate backbone depth. It also differs from the separate 13.22 s video-generation run. These are mean latencies without tail distributions or uncertainty for the small PDMS differences.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| NAVSIM v1 adaptive planning and latency Navtest, front camera without LiDAR; one trajectory per attempted exit; threshold 90; A100 80GB, batch one. | 90.79 PDMS (90.8 rounded); 170 ms; 94.1% exit by B15. PDMS ↑; mean end-to-end latency ↓ | Fixed B15: 90.62 / 190 ms; full path: 85.82 / 320 ms. The gain over B15 is 0.17 points with about 10% lower mean latency. The full path additionally decodes once; the 47% reduction is not an isolated depth-only comparison. e-v1e-routing-resultse-latency |
| NAVSIM v1 auxiliary proposal planning Fixed B22, 64 proposals, separate decoder and pseudo-expert coverage training; no adaptive routing. | 92.6 PDMS ↑ | Table 2 lists DriveVA at 90.5. The higher score changes decoder, supervision and proposal budget; it cannot be credited to adaptive exits or paired with 170 ms. e-v1e-auxiliary |
| NAVSIM v2 planning Official navtest EPDMS protocol; single front camera, no LiDAR. | 89.9 EPDMS ↑ | DriveVLA-W0 86.1; Epona 85.1; human reference 90.3. Best among the listed world-model planners. V2 uses different components and two-stage aggregation, so its scale is not interchangeable with v1. e-v2 |
| NAVSIM-to-nuScenes zero-shot transfer No nuScenes fine-tuning; averages over 1, 2 and 3 seconds. | 0.88 m; 0.08% L2 displacement ↓; collision rate ↓ | DriveVA: 0.84 m; 0.06%. Transfer is competitive but does not lead these two averages; it is not evidence of physical closed-loop driving. e-transfere-protocol |
| Joint adaptation ablation NAVSIM v1 single-trajectory comparison. | Joint LoRA 90.62; separate LoRA plus cached features 84.95. PDMS ↑ | Frozen Wan 84.20; full tuning 90.64; ViT-L 88.88. Joint adaptation adds 5.67 points over caching. The 0.02 full-tuning difference has no uncertainty estimate here. e-ablation |
| Offline scorer selection Fixed candidate pool spanning 12,146 scenes, separate from adaptive inference. | 91.2% / 94.4%; 51 cases with at least a 50-point gap. Exact / within-five-point selection ↑; large-gap errors ↓ | More than 95% of scenes contain all-perfect, all-zero or top-tied candidates. Strong selection accuracy under saturation does not certify safe early stopping on a shifted online pool. e-scorere-safety |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 2. Different depths solve overlapping, but nonidentical, sets of scenes. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read a row and column as two fixed exits. Appendix G defines each high-quality set using evaluated PDMS of at least 90 and averages intersection-over-union over ten aligned runs. The diagonal is 1.00 because a set overlaps itself; it is not a perfect planning score. Off-diagonal cells compare which scenes qualify at both exits relative to their union. B9 and B15 have the largest overlap, 0.82, while B15 and B30 overlap by 0.73. The legend begins at 0.65, so the lightest cells still indicate substantial sharing. This diagnostic complements the average fixed-exit scores rather than replacing them. e-overlape-depthe-routing-results
What it supports. B15 is the strongest fixed exit on average, yet its high-quality set does not exhaust the cases solved elsewhere. Reader interpretation: this supplies a reason to preserve deeper fallback and test input-dependent routing. It does not say that every deeper computation improves the currently available plan.
Where the evidence stops. Jaccard overlap measures agreement between thresholded sets, not scorer calibration, improvement magnitude or driving safety. Complementarity creates an opportunity; the learned verifier must still identify when further computation is worthwhile. The measured routing sweep provides the separate policy-level evidence.
Table 6. Joint LoRA adaptation retains nearly all the reported benefit of full tuning. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat the two halves as separate comparisons. On the left, the frozen Wan representation gives 84.20 PDMS, separate LoRA followed by cached features gives 84.95, and joint Wan LoRA reaches 90.62. Full tuning reports 90.64. These rows ask whether action learning can adapt the representation while the backbone and action heads are coupled. On the right, ViT-S, ViT-B and ViT-L give 83.91, 85.62 and 88.88, respectively. Those rows change the visual backbone; they are not paired combinations with the adaptation row beside them. The table concerns single-trajectory planning, with the auxiliary proposal comparisons supplied separately in Appendix I. e-ablatione-traininge-optim
What it supports. Joint LoRA improves on separately cached features by 5.67 PDMS, while full tuning adds only 0.02. Wan also exceeds the listed ViT-L result by 1.74 points. The evidence favors task adaptation and the chosen video representation in this setup, while leaving the statistical meaning of very small differences unresolved.
Where the evidence stops. These interventions do not isolate the video loss alone: backbone choice, pretraining and adaptation differ. No uncertainty is shown for the 0.02-point gap. Reader interpretation: a matched video-supervision ablation would be needed to attribute the gain specifically to predictive training.
7. Analysis & limitations
7.1 What the evidence leaves open
NAVSIM v1 uses non-reactive background actors; v2 approximates feedback with precomputed follow-up scenes. The authors explicitly deny a safety guarantee from the stopping threshold. e-protocole-v2e-safety
Source inconsistencies remain: Appendix F says the auxiliary B18 noise range is 0.14, while displayed extrema imply 0.16; Tables 16–17 show zero diagonal count means with nonzero standard deviations. Neither is silently repaired or used to strengthen a conclusion. e-noisee-pairwise
The six-output scorer and displayed v1 composition do not specify how the additional v2 components enter routing. Main result and routing tables provide no confidence intervals for the small adaptive gain. e-scorere-protocole-v2e-routing-results
7.2 Questions for discussion
- Does cross-depth routing outperform repeated B15 samples under matched decoding and scoring cost?
- Does the verifier preserve its false-accept rate when the scene distribution or candidate pool changes?
8. Reproducibility audit
8.1 Requirements and known gaps
Required resources include Wan2.2-TI2V-5B, DINOv2-Small, NAVSIM/OpenScene clips and evaluators. LoRA targets q/k/v/o with rank/alpha/dropout 32/64/0.05; its four-GPU recipe uses effective batch 80 for 80 epochs. Planner IL/GRPO use AdamW at 1e-4, 200/50 epochs and effective batches 1024/32. e-datae-methode-scorere-optim
Exact internal validation tokens, numerical actor loss weights, full scorer optimizer settings, software versions and coordination of the differing training schedules remain unspecified. Code is promised, not established as released. Timing replication must retain cached text and distinguish the 320 ms planning reference from the separate 13.22 s video-synthesis path. e-identitye-datae-optime-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: Separate cross-depth complementarity from extra sampling
Reader-proposed check, not performed: freeze the trained heads and scorer, then compare adaptive routing with repeated independent B15 samples on the same held-out scenes. Sweep B15 sample counts from one to six and report both matched candidate counts and measured latency on the same cached-text, batch-one implementation. Exclude future-video decoding from both planning paths. Evaluate learned selection and oracle selection separately, using identical observation inputs and paired scene-level score differences. Cross-depth complementarity is supported if different-depth candidates retain an oracle advantage and adaptive selection improves the quality–latency frontier. If repeated B15 sampling matches that frontier, the claimed value of varying representation depth is weakened. e-methode-depthe-overlape-routinge-latency
Check 2: Measure false acceptance on the actual routing pool
Reader-proposed check, not performed: retain the imitation-trained verifier and evaluate it on trajectories from the imitation and RL-refined actors, keeping evaluation scenes separate from threshold selection. Generate all six exit candidates offline for diagnosis, replay the exact accumulated-best stopping rule at thresholds 80, 90 and 95, and compare predicted versus evaluator quality. Report false acceptance below the threshold, 20- and 50-point selection gaps, and routing regret relative to the oracle best of six, stratified by exit and scores near the threshold. A marked increase after RL or a low offline selection error alongside frequent false early acceptance would challenge the transfer of the published scorer diagnostic to deployed routing. e-traininge-optime-scorere-routinge-routing-resultse-safety
8.3 Reading coverage
Visual audit: Visually inspected the title/authors/version on page 1; all five figures, Algorithm 1 and Tables 1–26 on the declared pages; and every final crop at its supplied pixel dimensions. Method, data, optimization, scorer, evaluation, profiling and transfer claims are supported by these pages, including uncropped appendix material. References on pages 8–9 were read as text. Figure 4 was checked against Eqs. (1)–(7), Algorithm 1 and Appendix D.4: strict versus inclusive threshold, ambiguous NO-path wiring and schematic LoRA/video branches are disclosed. Appendix F's range discrepancy and the directional-count tables' diagonal statistics remain unresolved in the base report. No separate supplement, code inspection or experiment reproduction is implied.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 10, 11, 12, 13, 14, 15, 16, 17. 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
- Title, authors, abstract and version statement
- 1 Introduction
- 2 Related Work
- 3 Motivating Analysis
- 4 Method, Algorithm 1 and Training
- 5 Experiments
- 6 Conclusion
- References
- Appendix A Roadmap
- Appendix B.1 NAVSIM and B.2 nuScenes protocols
- Appendix C.1–C.3 training clips, preprocessing and descriptions
- Appendix D.1–D.4 architecture, optimization and gradient isolation
- Appendix E auxiliary fixed-B22 model
- Appendix F video-noise analysis
- Appendix G.1–G.4 depth, overlap, directional counts and qualitative trajectories
- Appendix H.1–H.3 scorer architecture and diagnostics
- Appendix I ablations
- Appendix J routing
- Appendix K profiling
- Appendix L transfer
- Appendix M failure analysis
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The extraction limitation was addressed by inspecting original PDF pages 1–7 and 10–17 and all five final crops; reference-only pages 8–9 were read as text.
- Separate supplemental material availability has not been fully verified.
- Code and external linked resources were not inspected; no experiments were reproduced.
- Identity scope: the supplied title page verifies arXiv:2608.06008v1, 6 August 2026, with all three catalog authors. No title/author discrepancy or different edition was observed; no other revision was supplied.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title block, arXiv margin and abstract
The exact catalog title and all three authors match: Sining Ang, Yuguang Yang, Yan Wang. The artifact identifies arXiv:2608.06008v1 [cs.RO], 6 August 2026, and credits AIR, Tsinghua University; USTC Department of Automation; and Beihang School of Electronic Information Engineering. Code is promised for future release.
Go to primary source ↓e-interfacePDF p. 2, Figure 1 and caption
Adaptive-WAM uses trajectory exits and quality scoring; future-video prediction supervises training but is unnecessary for deployed planning.
Go to primary source ↓e-depthPDF p. 3, Section 3 and Table 1; p. 13, Appendix G.1, Table 14
Identical single-trajectory head schedules and validation-best selection over ten runs give B15 86.56 IL / 90.62 RL, versus B30 80.71 / 85.82. No PDMS confidence intervals accompany these table entries.
Go to primary source ↓e-noisePDF p. 3, Section 3, video timestep analysis; p. 13, Appendix F, Table 13
Single-trajectory IL scores span 0.13 at B15 and 0.15 at B18 across indices 1, 9, 17, 25, 32. The auxiliary B18 row ranges from 92.43 to 92.59, although adjacent prose calls its range 0.14.
Go to primary source ↓e-overlapPDF p. 3, Figure 2; p. 13, Appendix G.2, Eq. (19); p. 14, Table 15
High-quality scene sets use PDMS at least 90. Mean Jaccard overlap over ten aligned runs ranges from 0.69 to 0.82 off the diagonal; B9/B15 is 0.82 and B15/B30 is 0.73.
Go to primary source ↓e-pairwisePDF p. 4, Figure 3; p. 14, Appendix G.3, Eq. (20); p. 15, Tables 16–17
Directional counts concern score advantages of at least 50 points. Post-RL B15 over B30 is 598.64 and the reverse is 422.41. Diagonal means are zero but the displayed standard deviations are nonzero; the paper does not resolve that feature of the paired-run calculation.
Go to primary source ↓e-methodPDF p. 4, Section 4, Eqs. (1)–(2); p. 11, Appendices D.1–D.2
A Wan2.2-TI2V-5B conditional prefix uses the current front image and observation-derived description at video index 17. Scheduler noise initializes the future latent portion. Independent projected heads at blocks 5, 9, 15, 18, 22, 30 each perform five DDIM updates to output eight ego poses.
Go to primary source ↓e-routingPDF p. 4, Section 4, Eqs. (3)–(7); p. 5, Algorithm 1
The scorer takes an image and trajectory, predicts metric components and maintains the best accumulated predicted score. The controller stops at score greater than or equal to the validation-selected threshold, or at B30, and returns the best accumulated trajectory. Backbone prefixes and scores are reused.
Go to primary source ↓e-architecturePDF p. 5, Figure 4 and caption, Algorithm 1 line 9; p. 12, Appendix D.4
Figure 4 shows a strict greater-than branch, whereas Algorithm 1 uses greater-than-or-equal. The caption labels future-video generation as training supervision and repeated heads/LoRA locations as schematic; Appendix D.4 places LoRA throughout Wan.
Go to primary source ↓e-dataPDF p. 5, Training; pp. 10–11, Appendices C.1–C.3, Table 8
Temporal filtering retains 82,555 of 103,288 candidates as nine-frame clips at 2 Hz, removing 20,733. Wan images are 1280 by 704; scorer images are 832 by 480. Descriptions use anchor-time or historical information and cached T5 embeddings. Corpus counts do not specify the internal validation partition.
Go to primary source ↓e-trainingPDF pp. 5–6, Training, Eqs. (8)–(11); p. 11, Appendix D.3
Joint video and trajectory losses update Wan LoRA and the projected trajectory heads. Scorer updates alternate with actor updates, using detached trajectories and equal-weight soft-target BCE for six components. Planner-only DiffGRPO freezes Wan and the scorer.
Go to primary source ↓e-optimPDF pp. 12–13, Appendix D.4, Tables 10–11 and GRPO paragraph
Wan LoRA targets q/k/v/o with rank 32, alpha 64, dropout 0.05; its four-GPU recipe uses effective batch 80 for 80 epochs. Planner IL/GRPO schedules are 200/50 epochs with AdamW at 1e-4 and effective batches 1024/32. GRPO groups contain eight samples, with discount 0.6 and behavior-cloning weight 0.1. Numerical actor loss weights, full scorer optimization, software versions and exact schedule coordination are not supplied.
Go to primary source ↓e-scorerPDF p. 14, Appendices H.1–H.3, Tables 18–20; p. 17, Appendix M
DINOv2-Small uses image and flattened trajectory, without ego state or route command. Six heads predict NC, DAC, DDC, TTC, EP and comfort. On a separate 12,146-scene pool, exact/within-five-point selection is 91.2%/94.4%; 51 and 69 scenes have at least 50- and 20-point gaps. DINO-Small selected-candidate score is 92.59 versus Wan-B22 92.62, not end-to-end planner PDMS.
Go to primary source ↓e-v1PDF p. 6, Table 2, NAVSIM v1 navtest
Adaptive single-trajectory PDMS is 90.8, fixed-B22 64-proposal PDMS is 92.6, and listed DriveVA PDMS is 90.5. Adaptive NC/DAC/TTC/comfort/EP are 98.6/97.9/95.6/100/85.1; DriveVA values are 99.2/97.5/98.7/100/83.5. C and CL distinguish camera and camera-plus-LiDAR.
Go to primary source ↓e-v2PDF p. 6, Table 3; p. 10, Appendix B.1, Eqs. (13)–(14)
NAVSIM v2 navtest reports 89.9 EPDMS for Adaptive-WAM, 86.1 for DriveVLA-W0, 85.1 for Epona and 90.3 for the human agent. The metric uses additional components, human-reference penalty filtering and a two-stage follow-up aggregation.
Go to primary source ↓e-protocolPDF p. 10, Appendix B.1, Eq. (12)
NAVSIM v1 evaluates four-second trajectories in non-reactive simulation with recorded background futures and a controller for ego rollout. Its displayed normalized PDMS is NC times DAC times (5 EP + 5 TTC + 2 comfort)/12. V2 approximates closed-loop behavior without interactive simulation.
Go to primary source ↓e-auxiliaryPDF p. 12, Appendix E and Table 12; p. 13, continuation of Appendix E
The fixed-B22 comparator uses a separate learned-token decoder with 64 proposals and evaluator-filtered pseudo-expert coverage supervision. Training-time map and future occupancy construct targets; this is distinct from adaptive five-step single-trajectory decoding.
Go to primary source ↓e-routing-resultsPDF p. 7, Table 5; p. 15, Appendix J, Table 23
At thresholds 70/80/90/95, PDMS is 88.49/90.64/90.79/90.75 and mean latency is 112/143/170/284 ms. At 90, 94.1% stop by B15. Fixed B15 is 90.62 at 190 ms; full path is 85.82 at 320 ms.
Go to primary source ↓e-ablationPDF p. 7, Table 6; p. 15, Appendix I, Tables 21–22
Single-trajectory scores are 84.20 frozen Wan, 84.95 separate LoRA with caching, 90.62 joint LoRA and 90.64 full tuning. ViT-S/B/L give 83.91/85.62/88.88. The auxiliary 64-proposal columns are separate experiments.
Go to primary source ↓e-latencyPDF p. 7, Computation; pp. 15–16, Appendix K, Table 24
Planning latency uses one A100 80GB at batch one and includes VAE image encoding, conditional Wan prefix, attempted heads and scoring; cached text encoding is excluded. The full-depth reference additionally decodes once. The separate nine-frame, 40-step classifier-free video synthesis path takes 13.22 s, with 80 Wan forwards.
Go to primary source ↓e-transferPDF p. 7, Table 4; p. 10, Appendix B.2; pp. 16–17, Appendix L, Tables 25–26
NAVSIM-to-nuScenes zero-shot evaluation averages the 1/2/3-second metrics: Adaptive-WAM L2 0.35/0.71/1.58 m, average 0.88 m; collision 0.00/0.09/0.15%, average 0.08%. DriveVA averages 0.84 m and 0.06%.
Go to primary source ↓e-safetyPDF p. 17, Appendix M
The authors state that rare scorer failures matter, stricter thresholds cannot guarantee safety, and a passing quality threshold is not a formal safety certificate.
Go to primary source ↓8.5 Primary sources
Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features ↗
PDF · 9,599 extracted words
Source fingerprint
b8388c34a262ebbad0794c3da58afa1f81117d158728d3eea82f96e804094094