Driving in the Occupancy World: Vision-Centric 4D Occupancy Forecasting and Planning via World Models for Autonomous Driving
1. Paper overview
In one sentence: Drive-OccWorld couples an action-conditioned occupancy predictor with a cost-based planner, improving reported open-loop planning while leaving closed-loop safety and implementation reproducibility unresolved. e-architecturee-normalizatione-plannere-evaluatione-norm-ablatione-gaps
| At a glance | What to know |
|---|---|
| Research problem | Source description The authors seek to make a driving world model useful during planning, beyond generating training data or pretraining perception. Predicting occupied space and agent motion supplies future scene structure against which candidate trajectories can be evaluated. e-probleme-architecture |
| Core mechanism | Source description Semantic- and motion-conditioned memory combines voxel semantics, ego transforms and predicted agent flow before autoregressive forecasting. e-normalizatione-decoder |
| A key reported result | Open-loop trajectory planning — NoAvg: L2: 0.32, 0.75, 1.49; collision: 0.05, 0.17, 0.64; averages: 0.85 m / 0.29%. L2 error (m) / collision (%) at 1, 2 and 3 seconds; lower is better. nuScenes; predicted trajectories; Table 5 dagger block without the ego-status star. NoAvg evaluates each named timestamp. UniAD under the same block: average 1.03 m / 0.31%. Both averages improve, but collisions tie UniAD at 1 and 2 seconds. These are recorded-scene metrics. e-plan-resultse-evaluation |
| Reading caution | Reader analysis NuScenes open-loop results and selected qualitative generations do not establish closed-loop safety, distribution-shift robustness or calibrated counterfactual dynamics. No confidence intervals or repeated-seed variability are reported. e-evaluatione-qualitativee-gaps |
Core contributions
- Source description
Semantic- and motion-conditioned memory combines voxel semantics, ego transforms and predicted agent flow before autoregressive forecasting. e-normalizatione-decoder
- Source description
A common action interface supports controllable generation, while a separate planner closes the prediction–trajectory loop through occupancy costs and learned BEV refinement. e-actionse-plannere-architecture
Figure 2. A forecast–plan feedback loop with distinct generation and planning modes. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read from left to right: the history encoder converts camera observations into BEV embeddings, and the memory queue modulates them using semantics and motion. The world decoder receives this memory plus an action embedding and produces future BEV features, occupancy and flow. In the upper-right planner, agent and road predictions supply safety costs; future BEV features supply a learned cost. Sampled trajectories enter the argmin, followed by the plan decoder. Finally, trace the trajectory’s return arrow to the switch. Section 3.1 supplies the temporal ordering omitted by this schematic: a selected trajectory at t+1 conditions the state forecast at t+2. e-architecturee-decodere-plannere-actionse-evaluation
What it supports. The architecture supports a dual-system reading: world prediction and trajectory choice have distinct modules and roles. The shared BEV representation connects them, but the planner still scores proposals and refines a selected trajectory. The switch also matters experimentally: controllable generation can receive external action conditions while discarding the planner.
Where the evidence stops. The feedback arrow represents predicted-state rollout, not demonstrated vehicle execution. The schematic suppresses time offsets; Eqs. (3)–(4) and the following paragraph determine the sequence. It does not establish an independent world-model rollout for every candidate.
2. Motivation
2.1 The problem and the proposed response
The authors seek to make a driving world model useful during planning, beyond generating training data or pretraining perception. Predicting occupied space and agent motion supplies future scene structure against which candidate trajectories can be evaluated. e-probleme-architecture
2.2 What this reading follows
Imagine choosing a driving trajectory after first predicting where vehicles, pedestrians and road surfaces will be. Drive-OccWorld approaches this through a compact bird’s-eye-view representation: historical camera features become future semantic occupancy and motion, and a planner scores candidate movements against that forecast. The selected trajectory then conditions another prediction. The key reading distinction is between this planning loop and controllable generation supplied with known actions. The six excerpts below follow the information flow, inspect the normalization mechanism, and separate forecasting gains from planning gains. Their strongest lesson is metric-specific: better occupied-space prediction, better instance association and fewer predicted collisions do not always improve together. e-architecturee-normalizatione-plannere-evaluatione-norm-ablatione-gaps
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded Dual-system architecture is supported by explicit world model W and separate planner P, including a switch that removes P for generation. Cost-based proposal selection and BEV refinement constitute Other mechanisms: actions are neither jointly decoded with future states by one predictor nor recovered through a stated inverse-dynamics model. This supports Outside quadrants while retaining the WAM and multiview-driving scope. e-architecturee-planner
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 Build a memory that distinguishes scene layout from motion
Begin with historical camera images, not a precomputed occupancy grid. The BEVFormer encoder brings multiview geometry into a shared bird’s-eye-view representation. Drive-OccWorld then uses predicted semantics to modulate normalized features: an argmax produces voxel labels, one-hot encoding and convolutions produce scale and shift, and the modulated representation returns to BEV form. Motion supplies two additional sources of conditioning. Ego-pose transforms describe the vehicle’s own rotation and translation, while backward centripetal flow points from a voxel to its instance’s previous center. These signals address different changes in the observed scene. Table 6 is therefore informative beyond its best row: ego conditioning most strongly improves occupancy, whereas agent conditioning improves the flow-based association metric. Combining them preserves occupancy gains while recovering association quality. e-decodere-normalizatione-norm-ablation
Figure 3. Predicted voxel semantics determine how normalized features are scaled and shifted. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the upper branch first: a BEV embedding passes through layer normalization and a channel-to-height rearrangement. The lower branch predicts voxel semantics. The text specifies an argmax, one-hot labels and convolutions; the two convolution outputs become the semantic scale gamma and shift beta. At the right, the normalized representation is multiplied by gamma, then beta is added, matching Eq. (5). A height-to-channel rearrangement restores the BEV layout. This makes semantics an input to feature modulation as well as a prediction target. The figure shows the semantic pathway; the accompanying text separately defines ego-motion and agent-flow conditioning. e-normalizatione-norm-ablatione-gaps
What it supports. The memory operation changes features according to predicted scene content. Motion conditioning extends the same affine-modulation idea with ego transforms and backward centripetal flow. Table 6 then provides a way to ask whether each type of information helps occupancy forecasting, instance association, or both; the schematic alone cannot answer that.
Where the evidence stops. Semantic predictions can be wrong, so this pathway is not ground-truth semantic input. The supplied PDF does not specify gradient handling through argmax or the complete ordering of combined semantic, ego and agent modulation.
5.2 Keep supplied actions separate from planner feedback
An action condition tells the world decoder which ego motion or intention to incorporate. The interface accepts planar velocity, steering-derived curvature, next-timestamp displacement and a forward/left/right command. Fourier features and learned projection bring these heterogeneous inputs into the decoder’s conditional cross-attention. The paper then uses two modes. For action-controllable generation, the planner is discarded; for planning, predicted occupancy feeds costs, the selected trajectory is refined, and that prediction becomes an action condition for the next state forecast. Section 3.1 fixes this ordering even though Figure 2 compresses the loop. Table 4 exposes why the distinction matters: supplying ground-truth trajectories improves planning metrics, but that row is labeled an upper bound. A reproduction must preserve the boundary between future information supplied for analysis and actions available during inference. e-actionse-architecturee-plannere-action-results
5.3 Follow the evidence from forecasting to trajectory choice
Reader interpretation: better occupancy forecasts make useful planning inputs plausible, but the forecasting table alone cannot establish the planning mechanism. Table 1 shows stronger future occupancy and flow measures alongside weaker current occupancy accuracy. Table 5 then demonstrates improved open-loop trajectory metrics within carefully separated protocols. Table 8 provides a more local link: removing explicit safety costs or BEV refinement worsens collision results, even when average L2 barely changes. Together these experiments support the implemented combination of prediction, scoring and refinement. They do not isolate all benefits of future occupancy from shared BEV features, additional supervision or ego-motion information. A controlled replacement of future occupancy with current occupancy would test a narrower causal claim. Closed-loop interaction and behavior under distribution shift remain further questions, beyond the supplied nuScenes evidence. e-inflatede-plan-resultse-cost-ablatione-evaluatione-gaps
5.4 Training and inference
During training
Occupancy uses cross-entropy, Lovász and binary occupancy losses; flow uses l1 supervision. Planning combines max-margin candidate supervision, l2 imitation and collision loss. The PDF does not specify training stages, frozen modules or loss weights. e-traininge-gaps
The authors suggest predicted-trajectory training improves forecasting through planning constraints and ego-motion-aware BEV features. e-action-results
During inference
Controllable generation discards the planner to avoid potential ego-status leakage. Planning forecasts a state, selects and refines a trajectory, and feeds it back as the next action condition. Section 3.1 explicitly uses the trajectory at t+1 to forecast state t+2. e-architecturee-planner
This loop rolls out predicted states and trajectories. The reported evaluation does not execute those actions in traffic or test feedback from newly observed consequences. e-architecturee-evaluatione-gaps
5.5 Implementation flow
- Encode history
A BEVFormer history encoder transforms multiview image features into bird’s-eye-view (BEV) embeddings held in a memory queue. e-decoder
- Modulate memory
A prediction head produces semantic labels through argmax; one-hot labels and convolutions yield normalization parameters. Ego rotation/translation and predicted agent flow supply motion-conditioned parameters. Backward centripetal flow points from a voxel at time t to its instance center at t−1. e-normalization
- Forecast under an action
Learnable BEV queries use deformable self-attention, temporal cross-attention to memory, action cross-attention and a feedforward network. Fourier-encoded actions are concatenated and projected. Channel-to-height heads decode future occupancy and flow. e-decodere-actions
- Score and refine trajectories
The planner sums collision/proximity penalties, off-road penalties and a learned BEV cost map. It minimizes cost over sampled candidates, then refines a trajectory-and-command query by cross-attending to future BEV features and applying MLPs. e-planner
6. Experiments & results
Drive-OccWorld turns historical camera images into future semantic occupancy and flow, then uses those predictions to score and refine ego trajectories. Its central contribution is an inference-time feedback loop between a BEV world model and an occupancy-based planner. Reported nuScenes improvements concern forecasting and open-loop planning; they do not establish deployed driving safety.
6.1 Read the original evidence
Table 1. Future forecasting improves while current occupancy accuracy remains below Cam4DOcc. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the column subscripts before comparing rows. Current mIoU evaluates time zero, future mIoU evaluates predicted states, and the tilde marks the timestamp-weighted measure. VPQ assesses instance association through predicted flow; its starred variant uses center clustering and must remain separate. Then compare Cam4DOcc with Drive-OccWorld P: the P denotes predicted-trajectory conditioning, while A denotes ground-truth action conditioning. These are inflated movable-object occupancy labels derived from boxes, rather than the fine-grained background-inclusive labels of Table 2. The retained SPC footnote identifies its component methods; dashes and the word failed remain source omissions, not zero-valued results. e-inflatede-evaluatione-finee-norm-ablation
What it supports. Drive-OccWorld P raises future mIoU from 26.8 to 28.8 and VPQ from 21.4 to 23.3, gains of 2.0 and 1.9 score points. Current mIoU falls from Cam4DOcc’s 31.3 to 29.6. The evidence therefore favors future forecasting without showing that every perception measure improves.
Where the evidence stops. Table 1 identifies nuScenes validation but gives no numeric forecasting horizon. Section 4.4’s ablation configuration therefore cannot be assumed for this table, and the two horizons cannot be ordered. Starred VPQ and GT-conditioned rows also require separate comparisons.
Table 5. Protocol and ego-status markers determine which planning comparisons are meaningful. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the retained footnote: a dagger means NoAvg, a double dagger means TemAvg, and a star means the planner receives ego status. NoAvg evaluates the specified time; TemAvg averages performance from 0.5 seconds through that time. Within each block, smaller L2 distances and collision percentages are better. The Avg columns summarize the displayed horizons and do not erase the protocol distinction. First compare Drive-OccWorld P with UniAD in the unstarred dagger block, then repeat within the unstarred double-dagger block. Read the starred block separately because its input setting changes. A blank result in an older baseline remains unavailable. e-plan-resultse-evaluatione-gaps
What it supports. Under NoAvg, Drive-OccWorld P reports average L2 of 0.85 m versus UniAD’s 1.03 m, and collision of 0.29% versus 0.31%. Under unstarred TemAvg, the corresponding averages are 0.47 m/0.11% versus 0.69 m/0.12%. These are two separate favorable comparisons, not interchangeable measurements of one protocol.
Where the evidence stops. The evaluation is open loop on nuScenes. A lower recorded-scene collision rate is evidence about these predicted trajectories, not a demonstrated probability of safe deployment. Starred and unstarred rows must not be pooled.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Inflated occupancy and flow forecasting nuScenes validation; predicted-trajectory Drive-OccWorld P. Table 1 does not state a numeric forecast horizon. | 28.8 / 23.3 Future mIoU / VPQ; higher is better | Cam4DOcc: 26.8 / 21.4. Current mIoU: 29.6 versus 31.3. Future gains are 2.0 and 1.9 score points, while current occupancy accuracy is lower. Center-clustered VPQ* 33.2 is a separate association setting. e-inflatede-evaluation |
| Fine-grained occupancy forecasting nuScenes-Occupancy; predicted-trajectory model, 2-second future column. | 10.3 / 17.9 / 14.1 Movable / static / mean mIoU; higher is better | Cam4DOcc: 9.2 / 17.8 / 13.5. The movable-object gain exceeds the static-object gain; this benchmark includes background geometry absent from inflated movable-object labels. e-finee-evaluation |
| Open-loop trajectory planning — NoAvg nuScenes; predicted trajectories; Table 5 dagger block without the ego-status star. NoAvg evaluates each named timestamp. | L2: 0.32, 0.75, 1.49; collision: 0.05, 0.17, 0.64; averages: 0.85 m / 0.29%. L2 error (m) / collision (%) at 1, 2 and 3 seconds; lower is better | UniAD under the same block: average 1.03 m / 0.31%. Both averages improve, but collisions tie UniAD at 1 and 2 seconds. These are recorded-scene metrics. e-plan-resultse-evaluation |
| Open-loop trajectory planning — TemAvg nuScenes; Table 5 double-dagger block without the ego-status star. Each horizon averages from 0.5 seconds to that horizon. | 0.47 m / 0.11% Reported average L2 (m) / collision (%) | UniAD: 0.69 m / 0.12%. Compare within this protocol; these averages cannot substitute for NoAvg or the separate ego-status-starred results. e-plan-resultse-evaluation |
| Ground-truth-action planning upper bound nuScenes open-loop planning; Table 4 compares GT versus predicted trajectory conditioning. | GT: 0.56 m / 0.16%; predicted: 0.85 m / 0.29%. Average L2 (m) / collision (%) | GT future trajectory information improves both reported metrics. The GT row is explicitly an upper bound, not an available inference-time driving input. e-action-resultse-evaluation |
| Conditional-normalization ablation No action conditions; one historical plus current image set, two future forecasting timestamps; Table 6. | Baseline 26.4 / 33.5; ego-only 28.3 / 32.6; all three 28.3 / 34.5. Future mIoU at 1 second / VPQ* | Agent-only: 27.1 / 34.4. Ego conditioning helps occupancy yet harms instance association in isolation; the combined design balances the two. e-norm-ablatione-evaluation |
| Occupancy-cost and refinement ablation Table 8 planning ablation, reporting 0.5, 1 and 1.5 seconds. | Full: 0.28 m / 0.09%; no agent cost: 0.32 m / 0.16%; no refinement: 0.37 m / 0.20%. Reported average L2 (m) / collision (%) | Removing learned-volume cost gives 0.28 m / 0.14%. Safety costs matter even when average imitation error changes little. The shorter horizon prevents direct comparison with Table 5 averages. e-cost-ablation |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 6. Occupancy accuracy and instance association respond differently to motion conditioning. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each checkmark as enabling one normalization source, with the first row serving as the no-conditioning baseline and the final row enabling all three. Section 4.4 removes action conditions for this experiment, helping separate memory normalization from decoder action injection. Its default forecasting setup uses one historical and the current image set to predict two future timestamps. Compare future mIoU at one second with the final VPQ-star column: the former measures occupied-space semantics, while the latter also depends on flow-based instance association. The upward annotations emphasize improvements, but the unannotated decreases in VPQ-star are equally relevant to interpreting the mechanism. e-norm-ablatione-evaluatione-gapse-inflated
What it supports. Ego-only normalization changes future mIoU from 26.4 to 28.3 while VPQ* falls from 33.5 to 32.6. Agent-only conditioning gives 27.1 and 34.4; combining all three gives 28.3 and 34.5. The combined result supports complementary information, with different benefits for occupancy and instance association.
Where the evidence stops. Table 6 labels future mIoU at 1 second; Section 4.4 specifies one historical plus current input and two future timestamps. Table 1’s horizon is unspecified, so a relative-horizon comparison is unsupported. No seed variability is reported.
Table 8. Safety scoring and BEV refinement contribute beyond matching the reference trajectory. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each cross as removing one cost term or BEV refinement from the full bottom-row configuration. Agent cost penalizes overlap and excessive proximity to other road users; road cost penalizes leaving drivable space; volume cost is learned from future BEV features. BEV refinement updates the selected trajectory using those features. The result columns cover 0.5, 1 and 1.5 seconds, so their averages differ in scope from Table 5. Compare both L2 and collision: a configuration can match the full model’s average imitation error while becoming less safe under the paper’s collision metric. Inspect the per-horizon columns as well as their averages. e-cost-ablatione-plannere-evaluation
What it supports. Removing agent cost increases average collision from 0.09% to 0.16%; removing refinement raises it to 0.20% and L2 from 0.28 to 0.37 m. Without learned-volume cost, average L2 stays at 0.28 m while collision rises to 0.14%. This supports evaluating safety and imitation separately.
Where the evidence stops. The full model does not minimize every entry: removing volume cost lowers 1.5-second L2 from 0.46 to 0.44 m. The table tests component removal, not whether forecast occupancy beats current occupancy under fixed planner features.
7. Analysis & limitations
7.1 What the evidence leaves open
NuScenes open-loop results and selected qualitative generations do not establish closed-loop safety, distribution-shift robustness or calibrated counterfactual dynamics. No confidence intervals or repeated-seed variability are reported. e-evaluatione-qualitativee-gaps
Conditioning improvements are not uniform: Fourier cross-attention beats Fourier addition on future mIoU and VPQ*, but weighted mIoU falls from 28.0 to 27.8. Predicted-action versus GT-action forecasting also changes planning-related learning, so it does not isolate action quality alone. e-interfacee-action-results
7.2 Questions for discussion
- Would the occupancy-versus-association tradeoff survive matched-seed normalization ablations?
- How much planning gain remains when forecasts are replaced by current occupancy under identical ego-status inputs?
8. Reproducibility audit
8.1 Requirements and known gaps
Required ingredients include nuScenes images, bounding-box-derived inflated occupancy/instance labels, nuScenes-Occupancy voxel labels, ego transforms and action annotations. Keep GT-action generation, predicted-action planning, ego-status inputs and association algorithms separate. e-evaluatione-actionse-inflatede-plan-results
The PDF leaves optimizer, learning rate, epochs, batch size, initialization, frozen-module policy, numerical voxel resolution, trajectory-sampling count, cost thresholds/weights, loss weights, flow-clustering threshold and compute/runtime unspecified. Main-table forecasting history/horizons and full training/evaluation split details are not fully documented. Referenced methods are dependencies, not recovered implementation settings. e-gapse-traininge-plannere-inflated
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: Replicate the normalization tradeoff with controlled information
Reader-proposed check; not performed. Reproduce Table 6 with no action conditions, the same historical/current input sets, supervision and forecast horizon. Compare no conditional normalization, ego-only, agent-only and all three across matched seeds; add a shuffled-ego-transform negative control. Record both future mIoU and VPQ*, keeping the center-clustering rule fixed. Audit the timestamp and origin of every ego transform so future ground truth is not silently introduced. The source pattern predicts that ego-only improves occupancy more than association, while agent conditioning recovers VPQ*. If the gains disappear across seeds or persist with shuffled transforms, the claimed motion-specific explanation would be weakened. Missing optimization and clustering settings must be documented as reconstruction choices. e-normalizatione-norm-ablatione-evaluatione-gaps
Check 2: Isolate future occupancy in the planner’s safety costs
Reader-proposed check; not performed. Freeze the world-model features, learned-volume head, candidate trajectories and refinement module. Change only the occupancy supplied to explicit agent/road safety scoring: predicted future occupancy versus current occupancy reused at future times. Add an agent-cost-disabled control and a clearly labeled ground-truth-future-occupancy oracle. Keep navigation commands and ego-status inputs identical. Evaluate L2, collision and off-road incidence, reporting NoAvg and TemAvg separately at matching horizons. If forecast occupancy does not reduce collision relative to reused current occupancy, the specific value of explicit future safety scoring is unsupported in this controlled setting. This would not, by itself, rule out benefits carried by the frozen future BEV features. e-plannere-cost-ablatione-plan-resultse-evaluation
8.3 Reading coverage
Visual audit: All nine supplied PDF pages were rendered and visually inspected. Page 1 verifies title, author order, affiliations and AAAI-25 edition. Page 2 supplies the overview and introduction; page 3 supplies Figure 2 and rollout equations; page 4 supplies Figure 3, normalization, decoder, losses and action interface; page 5 supplies costs, refinement, planning losses, evaluation definitions and Tables 1–4; page 6 supplies Figures 4–5, action-setting definitions and result interpretation; page 7 supplies Tables 5–8, protocols, ablation settings and conclusion; pages 8–9 contain acknowledgments/references and confirm that no appendix follows. All six final original crops were separately viewed and found legible. Figure 3 multiplication/addition agrees with Eq. (5); Figure 2’s switch and return arrow were checked against Section 3.1, which supplies the temporal offsets. The Table 5 protocol footnote is intentionally retained. Separate supplements, code and external reference contents are outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9. 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
- Abstract
- 1 Introduction
- 2 Related Works
- 3 Method
- 3.1 Preliminary
- 3.2 4D Forecasting with World Model
- 3.3 Action-Controllable Generation
- 3.4 End-to-End Planning with World Model
- 4 Experiments
- 4.1 Setup
- 4.2 Main Results of 4D Occupancy Forecasting
- 4.3 End-to-end Planning with Drive-OccWorld
- 4.4 Ablation Study
- 5 Conclusion
- Acknowledgments
- References
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 four text chunks and all nine PDF pages were read. Original Figures 1–5 and Tables 1–8 were visually inspected, resolving the extraction-only figure omission.
- Identity/version: the title and all eight authors match the catalog. This is the AAAI-25 proceedings edition, printed pages 9327–9335. No revision identifier/history is printed; differences from other editions cannot be established.
- No appendix is present in this PDF; separate supplements were not supplied. Code and project-page contents were not inspected, and no experiment was reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1 (printed p. 9327), title, author block, affiliations and conference header
The exact catalog title and eight authors appear in the same order. Affiliations are Zhejiang University and Huawei Technologies. The header identifies AAAI-25; copyright is 2025. No revision identifier/history is printed.
Go to primary source ↓e-problemPDF pp. 1–2 (9327–9328), Abstract, Introduction and Figure 1
The proposed occupancy world model supports trajectory planning, beyond world-model data generation and visual-encoder pretraining.
Go to primary source ↓e-architecturePDF p. 3 (9329), Figure 2, Section 3.1, Eqs. (3)–(4) and following paragraph
History encoder, memory queue and world decoder predict states; a separate occupancy-based planner selects trajectories. Generation discards the planner. In planning, trajectory t+1 conditions state t+2.
Go to primary source ↓e-normalizationPDF p. 4 (9330), Figure 3, Section 3.2 and Eq. (5)
Non-affine layer normalization receives semantic/motion-derived scale and shift. Semantic labels use argmax, one-hot encoding and convolutions. Ego transforms encode rotation/translation; predicted backward centripetal flow encodes agent motion. Figure 3 multiplies by gamma, then adds beta.
Go to primary source ↓e-decoderPDF pp. 3–4 (9329–9330), Section 3.2, History Encoder and Future Forecasting with World Decoder; PDF p. 8 (9334), Li et al. 2022 BEVFormer reference
BEVFormer supplies visual BEV features. Queries undergo deformable self-attention, temporal and conditional cross-attention, and a feedforward network. Channel-to-height heads predict semantic occupancy and backward centripetal flow.
Go to primary source ↓e-actionsPDF p. 4 (9330), Section 3.3, Diverse Action Conditions and Unified Conditioning Interface
Velocity uses x/y speeds in m/s; steering feedback becomes curvature in inverse meters; trajectory uses next-timestamp displacement in meters; commands are forward, left or right. Fourier embeddings are concatenated and projected for cross-attention.
Go to primary source ↓e-plannerPDF p. 5 (9331), Section 3.4, Occupancy-based Cost Function and BEV Refinement
The planner sums agent-safety, road-safety and learned-volume costs, minimizes over sampled proposals, then refines a trajectory/command query through future-BEV cross-attention and MLPs.
Go to primary source ↓e-trainingPDF p. 4 (9330), Section 3.2 training paragraph; PDF p. 5 (9331), Section 3.4 planning-loss paragraph
Occupancy uses cross-entropy, Lovász and binary occupancy losses; flow uses l1. Planning uses max-margin, l2 imitation and collision losses. No optimizer, loss weights, training schedule or frozen-module settings are supplied.
Go to primary source ↓e-evaluationPDF p. 5 (9331), Section 4.1; PDF p. 7 (9333), Sections 4.3–4.4
Inflated occupancy uses bounding-box-derived labels; fine-grained occupancy uses nuScenes-Occupancy. Planning is open loop. NoAvg measures the named timestamp; TemAvg averages from 0.5 seconds to it. Default forecasting ablations use one historical plus current image set and two future timestamps.
Go to primary source ↓e-inflatedPDF p. 5 (9331), Table 1, Cam4DOcc and Drive-OccWorld rows; PDF p. 6 (9332), Section 4.2 opening and inflated-forecasting paragraphs
On nuScenes validation, predicted-trajectory Drive-OccWorld has future mIoU/VPQ 28.8/23.3 versus Cam4DOcc 26.8/21.4. Current mIoU is 29.6 versus 31.3. VPQ* 33.2 uses center clustering. A denotes GT actions; P denotes predicted trajectories. Table 1 prints no numeric forecasting horizon.
Go to primary source ↓e-finePDF p. 5 (9331), Table 2, GMO/GSO columns and caption; PDF p. 6 (9332), Section 4.2 fine-grained paragraph
At 2 seconds, Drive-OccWorld P movable/static/mean mIoU is 10.3/17.9/14.1 versus Cam4DOcc 9.2/17.8/13.5. Timestamp-weighted movable mIoU is 10.8 versus 9.7; current movable mIoU is 12.6 versus 11.0.
Go to primary source ↓e-action-resultsPDF p. 5 (9331), Tables 3–4; PDF p. 6 (9332), Section 4.2 Controllability
Table 3 future mIoU at 1 second: no conditions 26.4, GT trajectory 27.6, all four conditions 27.6, predicted trajectory 27.9. Table 4 GT-conditioned average L2/collision is 0.56 m/0.16%, versus predicted 0.85 m/0.29%; GT is labeled an upper bound. The proposed forecasting explanation is qualified as a possible effect of planning constraints.
Go to primary source ↓e-qualitativePDF p. 6 (9332), Figures 4–5 and captions
Figure 4 displays camera inputs, occupancy sequences and flow examples. Figure 5 shows different occupancy sequences under forward/left commands and two trajectories, without a quantitative counterfactual-response test.
Go to primary source ↓e-plan-resultsPDF p. 7 (9333), Table 5, protocol footnote and Section 4.3
NoAvg Drive-OccWorld P L2 is 0.32/0.75/1.49 m and collision 0.05/0.17/0.64% at 1/2/3 seconds; averages 0.85 m/0.29%, versus UniAD 1.03 m/0.31%. Unstarred TemAvg averages 0.47 m/0.11%, versus UniAD 0.69 m/0.12%. Starred TemAvg averages 0.32 m/0.29%; the star marks additional ego-status input.
Go to primary source ↓e-norm-ablationPDF p. 7 (9333), Table 6, future-mIoU header and semantic/ego/agent rows; Section 4.4 opening and Conditional Normalization paragraphs
Table 6 labels future mIoU at 1 second. Section 4.4 specifies one historical plus current input and two future forecasting timestamps. With action conditions removed, baseline future mIoU/VPQ* is 26.4/33.5; semantic-only 26.6/33.2; ego-only 28.3/32.6; agent-only 27.1/34.4; all three 28.3/34.5. Gains are metric-specific.
Go to primary source ↓e-interfacePDF p. 7 (9333), Table 7; Section 4.4 Action Conditioning Interface
Fourier plus addition gives future mIoU/weighted mIoU/VPQ* 27.4/28.0/34.2; Fourier plus cross-attention 27.6/27.8/35.0; cross-attention without Fourier 27.1/27.4/33.9. The preferred interface does not win every metric.
Go to primary source ↓e-cost-ablationPDF p. 7 (9333), Table 8; Section 4.4 Occupancy-based Costs
At 0.5/1/1.5-second horizons, full-model average L2/collision is 0.28 m/0.09%; no agent cost 0.32 m/0.16%; no road cost 0.29 m/0.11%; no volume cost 0.28 m/0.14%; no refinement 0.37 m/0.20%. No-volume 1.5-second L2 is 0.44 versus full-model 0.46 m.
Go to primary source ↓e-gapsPDF pp. 3–7 (9329–9333), complete Sections 3–5; PDF pp. 8–9 (9334–9335), Acknowledgments and References
The PDF supplies components, losses, tasks and tables but no implementation appendix, optimizer schedule, numerical voxel configuration, hardware/runtime, loss weights, sampling count or uncertainty estimates. Experiments use nuScenes; no closed-loop deployment study or dedicated limitations section is present.
Go to primary source ↓8.5 Primary sources
Driving in the Occupancy World: Vision-Centric 4D Occupancy Forecasting and Planning via World Models for Autonomous Driving ↗
PDF · 6,927 extracted words
Source fingerprint
075e829360df8283d6f882cbfb4fdedef8b3f48c001b645fb4d9aa40ce669685