PAPER REPORTENAll readings ↗

Pseudo-Simulation for Autonomous Driving

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Wei Cao; Marcel Hallgarten; Tianyu Li; Daniel Dauner; Xunjiang Gu; Caojun Wang; Yakov Miron; Marco Aiello; Hongyang Li; Igor Gilitschenski; Boris Ivanovic; Marco Pavone; Andreas Geiger; Kashyap Chitta

Affiliations: University of Tübingen, Tübingen AI Center; NVIDIA Research; Robert Bosch GmbH; OpenDriveLab at Shanghai Innovation Institute; University of Stuttgart; University of Toronto; Vector Institute; Stanford University

Source: CoRL 2025 · ref-361c60313b366d2109f0 ↗ · Catalog record

Reading: 413 / 558 · 6 original figures & tables · ~18 min ·

1. Paper overview

In one sentence: Pre-rendered future observations and endpoint-weighted scoring make driving evaluation sensitive to local deviations while retaining parallel planner queries, at the cost of incomplete interaction and per-scene reconstruction. e-probleme-flowe-aggregatione-correlatione-domaine-limitations

At a glanceWhat to know
Research problem
Source description

Expert-aligned open-loop tests miss recovery from a planner's own drift, while interactive sensor simulation is sequential and costly. The proposed compromise precomputes plausible future viewpoints, then emphasizes those near each evaluated planner's simulated endpoint. Its target is a more informative evaluation protocol, including sensitivity to motion-history shortcuts, rather than a newly trained driving policy. e-probleme-flow

Core mechanism
Source description

NAVSIM v2 combines reactive BEV trajectory scoring with pre-rendered counterfactual observations and proximity-weighted two-stage aggregation. e-rolloute-renderinge-aggregation

A key reported resultAlignment with nuPlan closed-loop evaluation: r = 0.89; R² = 0.8 (rounded).

Pearson r and coefficient of determination R²; higher is better.. Reported 83 privileged planners; 244 initial and 4,164 synthetic observations; eight-second nuPlan CLS versus two four-second stages. EPDMS excludes TLC, LK and EC.

Single-stage four/eight-second baselines: r = 0.83; R² = 0.7 (rounded). Supports simulator-score alignment for this planner family and subset; it does not establish score calibration, sensor-planner closed-loop validity or real-world safety. No uncertainty interval is reported. e-correlatione-limitations

Reading caution
Source description

Validation stops at simulator correlation. Residual rendering artifacts, centerline-constrained rule-based traffic and human-exception filtering can hide important failures. Reconstruction costs about 1–2 hours per scene on unspecified hardware; weighting and aggregation remain empirical design choices. e-limitations

Core contributions

  • Source description

    NAVSIM v2 combines reactive BEV trajectory scoring with pre-rendered counterfactual observations and proximity-weighted two-stage aggregation. e-rolloute-renderinge-aggregation

  • Author claim

    The authors report stronger simulator correlation and introduce navhard to expose perturbation sensitivity and comfort trade-offs in existing planners. e-correlatione-leaderboard

Figure 1. Precomputed observations become planner-specific tests through endpoint weighting. Original paper, p. 2 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Begin with the orange initial observation (a). The green trajectory ends near several purple synthetic starting states; each purple state has its own teal Stage 2 trajectory. The camera views above are already rendered before the tested planner runs. Follow the right-hand formula: first average the Stage 2 scores using normalized weights, then multiply by the Stage 1 score. The nearby purple states receive larger illustrative weights than the distant state. Equation (2) specifies distance to the endpoint reached in Stage 1 simulation, clarifying the caption's shorthand of a planned endpoint. Thus the diagram connects separate evaluations through scoring, rather than showing continuous sensor feedback. e-flowe-rolloute-aggregatione-limitations

What it supports. The mechanism makes a fixed bank of future observations useful for different planners: each planner's first simulated endpoint determines which alternatives matter most. A failure near that endpoint depresses the combined score, while a distant synthetic failure contributes little. This tests local recovery sensitivity without generating new camera observations online.

Where the evidence stops. The drawn weights are schematic, not measured success rates. Gaussian proximity is an empirical importance rule, not a demonstrated probability model for actual future states; no physical driving or continuous closed-loop sensor rollout is shown.

2. Motivation

2.1 The problem and the proposed response

Source description

Expert-aligned open-loop tests miss recovery from a planner's own drift, while interactive sensor simulation is sequential and costly. The proposed compromise precomputes plausible future viewpoints, then emphasizes those near each evaluated planner's simulated endpoint. Its target is a more informative evaluation protocol, including sensitivity to motion-history shortcuts, rather than a newly trained driving policy. e-probleme-flow

2.2 What this reading follows

A planner can look reliable when every test begins on the human driver's trajectory yet struggle after drifting away from it. Pseudo-simulation addresses this evaluation gap with two sets of inputs: a recorded observation and a precomputed bank of plausible future observations. It scores the planner independently at each input, then emphasizes synthetic starts close to the endpoint reached by its initial plan. The resulting NAVSIM v2 benchmark links open-loop efficiency with some recovery-sensitive testing. Read the evidence in three parts: how observations are connected, whether scores track a closed-loop simulator, and whether synthetic rendering itself explains the observed failures. e-probleme-flowe-aggregatione-correlatione-domaine-limitations

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 dimensionRecorded classification
Major categoryBenchmarks & simulators
ArchitectureNot applicable
Prediction paradigmNot applicable
QuadrantNot applicable

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

The recorded benchmark/simulator classification is supported. Scene reconstruction generates evaluation inputs; independent planners produce trajectories, and the evaluator links them through scores. This is not evidence of one shared world-action model, joint future/action prediction or inverse dynamics. Architecture, prediction paradigm and quadrant are appropriately not applicable to this paper's contribution. e-flowe-renderinge-aggregation

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

InputsOutputs
  • Planner: multi-view camera images, ego velocity/motion history, and left/straight/right command.
  • Evaluator: recorded scenes, expert trajectories, annotations and pre-generated synthetic observations.
  • Planner: four-second local-coordinate waypoint trajectory.
  • Evaluator: Stage 1 and Stage 2 EPDMS subscores and combined pseudo-simulation score.

4.2 Equations and their role

EPDMS=(mMpenfilterm(agent,human))mMavgwmfilterm(agent,human)mMavgwm\mathrm{EPDMS}=\left(\prod_{m\in\mathcal M_{\mathrm{pen}}}\mathrm{filter}_m(\mathrm{agent},\mathrm{human})\right)\frac{\sum_{m\in\mathcal M_{\mathrm{avg}}}w_m\mathrm{filter}_m(\mathrm{agent},\mathrm{human})}{\sum_{m\in\mathcal M_{\mathrm{avg}}}w_m}
Equation (1): penalty metrics are NC, DAC, DDC and TLC; averaged metrics are TTC, EP, HC, LK and EC. Table 1 defines their names/ranges. Weights are 5 for progress/TTC and 2 for lane keeping/history comfort/extended comfort. The filter ignores violations also committed by the human expert. e-metric
scombined=s1s2,s2=iw^is2i,w^i=wiiwi,wi=exp ⁣(xix^22σ2)s_{\mathrm{combined}}=s_1s_2,\qquad s_2=\sum_i\hat w^{i}s_2^{i},\qquad \hat w^{i}=\frac{w^{i}}{\sum_i w^{i}},\qquad w^{i}=\exp\!\left(-\frac{\|x^{i}-\hat x\|^2}{2\sigma^2}\right)
Equation (2): s₁ is the initial score, s₂ⁱ a synthetic-start score, xⁱ that start position, and x̂ the endpoint reached in Stage 1 simulation. Normalized weights favor nearby starts; σ² controls kernel variance, default 0.1. This is an empirical proximity rule, not a calibrated future-state probability. e-aggregatione-limitations

5. Method in detail

5.1 Generate the test bank before asking the planner what it would do

Reader analysis

The ordering is central. Stage 2 states are sampled around the expert's endpoint after four seconds, before any particular planner is evaluated. The pipeline matches each accepted position to a plausible human-derived heading and motion history, removes invalid states, and reconstructs camera observations with a single-traversal MTGS variant. Delaying the synthetic observation allows traffic to react and supplies a plausible preceding history. At evaluation time, the planner's real-observation trajectory is executed in BEV to obtain its own endpoint. That endpoint selects no new rendering; it changes the weights assigned to existing tests. As a reader interpretation, this is a conditional reuse strategy: one expensive observation bank can support many planner evaluations, but its coverage remains tied to states sampled around the expert. e-samplinge-renderinge-rolloute-aggregation

Table 1. EPDMS combines rule penalties with progress, risk and comfort subscores. Original paper, p. 4 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the table in two blocks. The upper rows are multiplicative penalties: no at-fault collision (NC), drivable-area compliance (DAC), driving-direction compliance (DDC) and traffic-light compliance (TLC). The lower block enters a weighted average: ego progress (EP), time to collision (TTC), lane keeping (LK), history comfort (HC) and extended comfort (EC). The middle column assigns weights of 5 to EP/TTC and 2 to the other averaged terms. The range column distinguishes continuous progress from discrete penalties, including half-credit values for NC and DDC. Equation (1) also applies human-violation filtering before combining the components. e-metrice-correlatione-limitations

What it supports. A zero-valued retained penalty can nullify the overall score, even when progress and comfort are high. Consequently, the final metric encodes explicit priorities rather than treating all driving attributes as interchangeable. Keeping the individual subscores visible is essential for interpreting why a planner receives its aggregate score.

Where the evidence stops. The correlation experiment omits TLC, LK and EC because nuPlan lacks those closed-loop metrics; the navhard leaderboard uses the fuller score. Human-expert exceptions can also suppress genuine safety failures, as the authors acknowledge.

5.2 Understand what the two products penalize

Source description

There are two different multiplications in the protocol. Within each segment, EPDMS multiplies rule-compliance terms by a weighted average of progress, risk and comfort measures. Across segments, the final score multiplies the first-stage score by the Gaussian-weighted average of second-stage scores. A nearby synthetic failure can therefore reduce the score even if the original trajectory performs well; conversely, strong recovery scores do not erase a failed first segment. The Gaussian uses the distance between a synthetic start and the endpoint actually reached by the simulated first plan. This mechanism approximates a continuation test without sending intermediate sensor feedback to the planner. The authors support their multiplication choice through correlation ablations, while acknowledging that the weighting and aggregation remain empirical rather than a principled probabilistic formulation. e-metrice-aggregatione-rolloute-ablatione-limitations

Figure 2. Future viewpoint changes alter both road geometry in the image and proximity to traffic. Original paper, p. 5 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Treat each quadrant as one scene. The orange border and orange pose mark the initial real observation; purple borders and pose arrows mark pre-generated alternatives. The connecting lines associate particular camera images with poses on the road diagram. Compare the road markings and nearby vehicles across views: the bus example makes the effect of longitudinal displacement especially visible. These are selected examples of the sampled observation bank, not a frame-by-frame rollout from a tested planner. Section 3.2 explains that sampling occurs around the expert's future endpoint and attaches a plausible human-derived motion history, although those histories are not plotted here. e-scenese-samplinge-renderinge-domaine-nvs

What it supports. The figure shows why testing new viewpoints is more than re-rendering the same image: the spatial relationship between ego, road and traffic changes. A planner that relies on expert-aligned appearance or history may therefore behave differently. The examples establish the kinds of perturbations supplied to evaluation, without quantifying recovery performance.

Where the evidence stops. These selected images cannot establish fidelity across all scenes or sampled positions. Low-quality reconstructions are filtered, and the precise semi-automatic visual-quality procedure is deferred to unavailable supplementary material. Quantitative rendering evidence is limited to Table 3's separate tests.

5.3 Separate evidence about the evaluator from evidence about a planner

Reader analysis

The experiments answer different questions. The nuPlan study asks whether pseudo-simulation scores correlate with closed-loop scores across privileged planners; it omits three EPDMS components to match available closed-loop metrics. The navhard leaderboard instead uses the fuller score and includes planners with different sensor privileges, exposing stage-specific failure modes. The rendering diagnostic then holds a real-trained LTF model fixed and compares real S1, synthetic S1 and synthetic S2. Its pattern is consistent with viewpoint/state deviations contributing more to the planning drop than replacing real images with rendered images at S1. As a reader interpretation, this supports the evaluation mechanism but does not completely disentangle rendering artifacts from difficult states. Neither the correlation study nor these diagnostics demonstrates a quantitative relationship to real-world vehicle safety. e-correlatione-leaderboarde-domaine-limitations

5.4 Training and inference

During training

Source description

Scene preparation optimizes a scene-graph Gaussian representation from one traversal. LiDAR registration and bundle adjustment initialize camera poses, followed by pose optimization during reconstruction. Sensor failures and low-quality reconstructions are filtered. The supplied body does not specify a complete optimization schedule. e-rendering

Reader analysis

The planner is an evaluated component, not jointly trained with this renderer. The fidelity diagnostic holds a real-data-trained LTF model fixed. No new planner loss, frozen-module schedule or joint action/video objective is presented. e-traininge-rendering

During inference

Reader analysis

All sensor observations already exist at evaluation time, so planner queries can be parallelized; only score aggregation depends on the first-stage endpoint. Low-level simulated trajectory execution must be distinguished from online planner feedback and physical vehicle execution. e-flowe-rolloute-aggregation

5.5 Implementation flow

  1. Execute a committed plan

    From the real observation, obtain a four-second trajectory. A bicycle model and LQR controller execute it at 10 Hz in BEV while IDM background vehicles react. The planner receives no new observation during this segment; simulated execution produces its score and endpoint. e-flowe-rollout

  2. Precompute plausible alternative starts

    Before evaluating any planner, sample around the expert's four-second endpoint: every 0.5 m laterally within ±2 m, and every 5 m longitudinally within reachability under ±4 m/s². Match human histories, rejecting differences above 1 m/s velocity, 1 m/s² acceleration or 20 degrees heading. Reject invalid penalty states and scenes with fewer than five valid synthetic observations. e-sampling

  3. Render and evaluate alternatives

    Single-traversal MTGS reconstruction supplies synthetic cameras at accepted states. The four-second delay accommodates background reactions and plausible histories. Independently query the planner and repeat BEV scoring at each synthetic start. e-samplinge-renderinge-rollout

  4. Connect stages through scoring

    Weight synthetic scores by proximity to the Stage 1 simulated endpoint, normalize the weights and multiply the resulting average by the Stage 1 score. Synthetic-state generation is planner-independent; its contribution to the final score is planner-dependent. e-aggregation

6. Experiments & results

Pseudo-simulation evaluates driving planners on recorded observations and a fixed bank of synthetic future observations. Gaussian proximity weights connect the two stages without online sensor rendering. NAVSIM v2 thereby tests responses to deviations from expert driving while permitting parallel evaluation. Its strongest evidence is improved correlation with nuPlan closed-loop scores; it does not establish real-world safety or a learned world-action architecture (e-flow, e-correlation, e-limitations).

6.1 Read the original evidence

Table 2. Stage-wise subscores reveal weaknesses hidden by the overall leaderboard order. Original paper, p. 7 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read horizontally to compare planners on the same metric and vertically within each paired row to compare original S1 with synthetic S2 observations. CV is constant velocity; MLP uses ego history; LTF is the image-only Latent TransFuser planner; PDM-C is privileged PDM-Closed. The rightmost column is visually separated and italicized. The bottom EPDMS row is the combined two-stage score, not an arithmetic average of the displayed row values. To understand the ranking, inspect both lane keeping (LK) and extended comfort (EC): a method can lead the combined score while still performing poorly on a particular driving attribute. e-leaderboarde-metrice-correlation

What it supports. PDM-C leads combined EPDMS with 51.3, followed by LTF at 23.1, MLP at 12.7 and CV at 10.9. LTF's lane-keeping score falls from 94.2 to 45.4 across stages. PDM-C's extended-comfort scores are only 36.0 and 25.4, illustrating the comfort trade-off emphasized by the authors.

Where the evidence stops. The planners have different input privileges, so the ranking is not a controlled architecture comparison. Scores are displayed on a 0–100 scale but are not physical-driving success percentages. This full-metric navhard test differs from the reduced-metric correlation experiment.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
Alignment with nuPlan closed-loop evaluation

Reported 83 privileged planners; 244 initial and 4,164 synthetic observations; eight-second nuPlan CLS versus two four-second stages. EPDMS excludes TLC, LK and EC.

r = 0.89; R² = 0.8 (rounded).

Pearson r and coefficient of determination R²; higher is better.

Single-stage four/eight-second baselines: r = 0.83; R² = 0.7 (rounded).

Supports simulator-score alignment for this planner family and subset; it does not establish score calibration, sensor-planner closed-loop validity or real-world safety. No uncertainty interval is reported. e-correlatione-limitations

Aggregation and synthetic-observation density ablations

The same correlation study varies stage count, Gaussian variance, aggregation and synthetic-view density.

Small variances (0.05/0.1) perform best; multiplication beats mean/hybrid aggregation; correlations exceed 0.85 at 25% density.

Pearson and Spearman correlations with CLS.

The authors state 13 planner calls versus 80 for an eight-second, 10 Hz closed-loop rollout, approximately sixfold fewer.

These are evaluation-query savings, not measured total-runtime savings. Reported observation totals and average density are inconsistent; preprocessing is excluded from this comparison. e-ablatione-countse-limitations

NAVSIM v2 navhard planner comparison

450 initial and 5,462 synthetic observations; full EPDMS; four planners with different input privileges.

PDM-C 51.3; LTF 23.1; ego-history MLP 12.7; constant velocity 10.9.

Combined EPDMS, displayed on the table's 0–100 scale; higher is better.

LTF lane keeping drops from 94.2 in S1 to 45.4 in S2; PDM-C extended comfort is 36.0/25.4.

Stage-wise diagnosis reveals robustness and comfort weaknesses. Privileged PDM-C and image-only LTF are not matched-input comparisons; scores are not physical-driving success percentages. e-leaderboarde-metric

Rendering fidelity and downstream sensitivity

Real-trained LTF on real S1, synthetic S1 and synthetic S2 navhard observations. Separate LPIPS test: eight scenes, alternating 10 Hz expert-trajectory train/test views.

mIoU/EPDMS: 46.0/62.3, 37.6/61.0, 36.9/44.2. Full-renderer LPIPS: 0.253.

BEV mIoU and planning EPDMS ↑; LPIPS ↓.

LPIPS is 0.354 for Street Gaussians and 0.322 for the variant labeled without pose optimization.

The fixed-pose rendering gap affects planning less than the subsequent state perturbation. This is evidence from one planner; the renderer comparison bundles several pose-related changes. e-domaine-nvs

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Figure 3(a–e). Two-stage evaluation improves correlation, and the diagnostic panels test its aggregation choices. Original paper, p. 6 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. In panel (a), each point represents a planner, with CLS on the horizontal axis and pseudo-simulation EPDMS on the vertical axis. Blue and orange distinguish rule-based and learning-based planners; the two axes use different ranges. Panels (b–e) separate Spearman rank correlation from Pearson linear correlation using two bar colors. Bold-underlined settings are the authors' defaults. Read panel (b) for two four-second stages versus single-stage alternatives, (c) for kernel variance, (d) for multiplication versus other stage combinations, and (e) for reducing the observation bank. The plotted variance is sigma squared, not a standard deviation or learned uncertainty estimate. e-correlatione-ablatione-countse-limitations

What it supports. The text reports Pearson r = 0.89 for two stages against 0.83 for single-stage baselines. Multiplication has the strongest correlation among the displayed stage aggregators, and reducing the observation bank to 25% still leaves both correlations above 0.85. These findings support the chosen evaluation design on the tested simulator subset.

Where the evidence stops. These are privileged-planner correlations with reduced EPDMS, without uncertainty intervals or real-world validation. The caption's 37 rule-based planners conflicts with the prose's listed counts; the later mean observation count also conflicts with the stated subset totals.

Table 3(a–b). Matched-stage rendering and later state perturbation have different downstream effects. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. In panel (a), first compare Real S1 with Syn. S1: this changes rendering while retaining the first-stage setting. Next compare Syn. S1 with Syn. S2 to examine perturbed synthetic observations. The left metric is BEV segmentation mIoU averaged over drivable area, walkway and vehicle classes; the right is planning EPDMS for the same real-trained LTF model. Panel (b) is a separate eight-scene rendering test, using alternating expert-trajectory views for training and evaluation. Its downward arrow means lower LPIPS is better. The full method adds the pose-related processing described in the text; the abbreviated middle-row label does not isolate every change. e-domaine-nvse-rendering

What it supports. Real-to-synthetic S1 reduces mIoU from 46.0 to 37.6 while EPDMS changes only from 62.3 to 61.0. Synthetic S2 reaches 36.9 mIoU but 44.2 EPDMS, consistent with additional state-shift sensitivity. The full renderer has LPIPS 0.253 versus 0.322 for the ablated variant and 0.354 for Street Gaussians.

Where the evidence stops. The Results prose mislabels the 46.0-to-37.6 comparison as S1 versus S2; the table clearly identifies real versus synthetic S1. One planner and eight rendering scenes do not establish universal fidelity, and the pose-processing comparison bundles changes.

7. Analysis & limitations

7.1 What the evidence leaves open

Source description

Validation stops at simulator correlation. Residual rendering artifacts, centerline-constrained rule-based traffic and human-exception filtering can hide important failures. Reconstruction costs about 1–2 hours per scene on unspecified hardware; weighting and aggregation remain empirical design choices. e-limitations

Reader analysis

Unresolved reporting inconsistencies: Figure 3 says 37 rule-based plus 46 learned planners, while the listed rule-based counts sum to 40. Also, 4,164/244 implies roughly 17.1 synthetic observations, not the stated mean of 12. Table 3(a) labels 46.0→37.6 as real→synthetic S1, although nearby prose calls it S1→S2; this report follows the table. e-countse-domain

7.2 Questions for discussion

  1. Does proximity weighting preserve useful rankings when a planner leaves the region sampled around the expert? (e-sampling, e-aggregation)
  2. Would the renderer-versus-perturbation diagnosis hold across several sensor-based planners and more interactive traffic? (e-domain, e-limitations)

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

Reproduction requires the exact nuPlan subsets, planner configurations, expert histories, camera/LiDAR calibration, reconstruction settings, reactive-traffic/controller configuration and EPDMS implementation. The supplied body defers subset selection, metric modifications and reconstruction-quality filtering to absent supplementary material; it supplies some sampling thresholds but not a complete implementation or software/hardware specification. e-rolloute-metrice-samplinge-renderinge-correlatione-leaderboarde-limitations

Reader analysis

Proposed checks: hold scenes and planner outputs fixed while varying aggregation; separately compare rendering-only and viewpoint-perturbation effects with fixed planners. Resolve planner/observation counts before interpreting efficiency or reproducing correlations. e-ablatione-countse-domaine-nvs

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: Test whether endpoint weighting generalizes beyond the configurations used to choose it

Reader-proposed experiment, not performed: first reconcile the listed planner counts and observation totals, then freeze a documented scene subset, synthetic bank, planner outputs and reduced EPDMS definition. Compare the default Gaussian variance 0.1 and multiplicative stage aggregation against uniform weighting and mean/hybrid stage aggregation. Select settings on one scene partition and evaluate Pearson r and Spearman rho against eight-second CLS on a held-out partition, reporting scene-bootstrap intervals and planner-family results. If the default gains disappear or reverse outside the selection partition, its apparent advantage would reflect subset dependence rather than robust evaluation design. e-correlatione-aggregatione-ablatione-counts

Check 2: Distinguish reconstruction error from sensitivity to perturbed states

Reader-proposed experiment, not performed: extend Table 3's three-condition comparison to several fixed sensor-based planners, keeping their training and scene set unchanged. For each scene, evaluate real S1, re-rendered S1 and synthetic S2 under both the full and ablated reconstruction pipelines. Use paired scene-level changes in EPDMS and BEV mIoU, and report LPIPS only where held-out real reference views exist. Track rejected scenes rather than silently changing the subset. If improved rendering largely restores S2 planning scores, a rendering explanation gains support; persistent S2 degradation across renderers and planners would strengthen the state-shift interpretation. e-domaine-nvse-renderinge-sampling

8.3 Reading coverage

Visual audit: Rendered and visually inspected the title/author page and every method, metric, numerical-result, ablation and limitation page supporting this bundle. Inspected Figures 1–3, Tables 1–3, Eqs. (1–2) and the unnumbered comparison on p. 2; viewed every final crop. Figure 2 retains its short caption because it contains the original color/pose legend. Figure 1's weighting diagram was checked against its caption and Eq. (2); the explanation uses the simulated endpoint defined by the equation. Table 3(a)'s row labels resolve the identified prose mislabel, but planner and observation count discrepancies remain unresolved. References on pp. 10–14 were read as text; they support no additional retained method or numerical claims. No appendix or separate supplement was supplied, and code and external resources were not inspected.

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 Work
  • 3 Pseudo-Simulation
  • 3.1 Stage 1: Initial Observations
  • 3.2 Stage 2: Synthetic Observations
  • 4.1 Closed-loop correlation study
  • 4.2 NAVSIM v2/navhard leaderboard
  • 4.3 Rendering fidelity and ablation
  • 5 Conclusion
  • Limitations and Future Work
  • Acknowledgments
  • References (PDF pp. 10–14)

Outside the original text pass

  • Identity: title and all 14 authors agree with the catalog. The inspected artifact is labeled CoRL 2025; no numbered revision or separate revision date is established, and no alternative edition was supplied for comparison.
  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • The preceding extraction limitation was addressed by inspecting PDF pages 1–9, all three numbered figures and all three numbered tables, including equations and captions.
  • Separate supplemental material availability has not been fully verified.
  • No appendix is present in the supplied PDF. Referenced supplementary metric modifications, subset/filtering details and reconstruction-quality filtering details were not supplied.
  • Code, external project resources and leaderboard services were not inspected; no experiments were reproduced. The paper's code-availability statement is not independently verified.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e-identityPDF p. 1, title block and conference footerInspect

The observed title is Pseudo-Simulation for Autonomous Driving. All 14 authors match the catalog in order. The footer identifies the 9th Conference on Robot Learning (CoRL 2025), Seoul, Korea; no numbered revision is shown.

Go to primary source ↓
e-problemPDF pp. 1–3, Abstract and Sections 1–2Inspect

The paper motivates scalable evaluation beyond expert-aligned open-loop observations and introduces NAVSIM v2 as an evaluation framework using pre-rendered perturbations.

Go to primary source ↓
e-flowPDF p. 2, Figure 1 and caption; p. 3, Section 3Inspect

Real and pre-generated synthetic observations are evaluated separately. Planners consume multi-view cameras, ego status/history and a left/straight/right command, outputting four-second local-frame waypoint trajectories.

Go to primary source ↓
e-rolloutPDF pp. 3–4, Section 3.1, BEV SimulationInspect

A kinematic bicycle model and LQR controller execute a committed four-second trajectory at 10 Hz, without planner feedback within that horizon. IDM provides reactive background vehicles represented by oriented boxes.

Go to primary source ↓
e-metricPDF p. 4, Section 3.1, Eq. (1), Table 1 and human-violation filtering paragraphInspect

EPDMS multiplies NC, DAC, DDC and TLC penalties by a weighted average of TTC, EP, HC, LK and EC. EP/TTC weights are 5; LK/HC/EC weights are 2. Violations also committed by the human expert are ignored.

Go to primary source ↓
e-samplingPDF pp. 4–5, Section 3.2, Start Point Sampling and Heading and History GenerationInspect

Sampling is centered on the expert endpoint after four seconds: lateral spacing 0.5 m within ±2 m, longitudinal spacing 5 m within reachability under ±4 m/s² acceleration. Human-history matching uses velocity, acceleration and heading thresholds of 1 m/s, 1 m/s² and 20 degrees. Invalid penalty states are rejected; scenes with fewer than five valid synthetic observations are discarded.

Go to primary source ↓
e-scenesPDF p. 5, Figure 2 and captionInspect

Four scene examples pair orange real observations with purple synthetic views and pose markers; the examples show changes in position and heading around traffic and intersections.

Go to primary source ↓
e-renderingPDF p. 5, Section 3.2, Neural Reconstruction and RenderingInspect

The MTGS variant uses single-traversal scene-graph reconstruction, LiDAR registration, bundle adjustment and camera-pose optimization. Sensor-failure scenes and low-quality reconstructions are filtered; the latter procedure is deferred to supplementary material.

Go to primary source ↓
e-aggregationPDF p. 5, Section 3.2, Score Aggregation and Eq. (2); p. 6, Section 4.1, weighting resultsInspect

Stage 2 scores are averaged with normalized Gaussian weights based on distance from each synthetic start position to the Stage 1 simulated endpoint. The final score multiplies Stage 1 and aggregated Stage 2 scores; the default kernel variance is 0.1.

Go to primary source ↓
e-correlationPDF p. 6, Figure 3(a–b), caption and Section 4.1, Benchmark/ResultsInspect

The reported correlation study uses 83 privileged planners, 244 initial and 4,164 synthetic observations, and reduced EPDMS without TLC/LK/EC. Against eight-second nuPlan CLS, two four-second stages yield Pearson r=0.89 and rounded R²=0.8, versus r=0.83 and rounded R²=0.7 for single-stage baselines.

Go to primary source ↓
e-ablationPDF p. 6, Figure 3(b–e) and variance discussion; p. 7, Section 4.1 continuationInspect

Two stages outperform single-stage settings; small Gaussian variances perform best; multiplication outperforms mean/hybrid aggregation. At 25% synthetic-observation density, correlations remain above 0.85. The efficiency paragraph states 13 versus 80 planner inferences per scenario.

Go to primary source ↓
e-countsPDF p. 6, Figure 3 caption and Section 4.1 planner list/subset counts; p. 7, Section 4.1 density paragraphInspect

Figure 3 states 37 rule-based plus 46 learned planners, totaling 83, but the prose lists 10+15+15 rule-based variants. The stated 4,164 synthetic observations over 244 initial observations also do not match the later claimed mean of 12 synthetic observations per initial observation; the PDF gives no reconciliation.

Go to primary source ↓
e-leaderboardPDF p. 7, Section 4.2 and Table 2, navhard leaderboardInspect

Navhard has 450 initial and 5,462 synthetic observations. Combined EPDMS is CV 10.9, MLP 12.7, LTF 23.1 and privileged PDM-C 51.3. LTF lane keeping is 94.2/45.4 for S1/S2; PDM-C extended comfort is 36.0/25.4.

Go to primary source ↓
e-domainPDF p. 8, Table 3(a), caption and Section 4.3, Benchmark/ResultsInspect

Real-trained LTF obtains mIoU/EPDMS of 46.0/62.3 on real S1, 37.6/61.0 on synthetic S1, and 36.9/44.2 on synthetic S2. mIoU averages drivable area, walkway and vehicle classes. The first Results paragraph mislabels the 46.0-to-37.6 comparison as S1 versus S2; the table labels it real versus synthetic S1.

Go to primary source ↓
e-nvsPDF p. 8, Table 3(b), caption and Section 4.3, Ablation StudyInspect

LPIPS on eight navhard scenes with alternating 10 Hz expert-trajectory train/test views is 0.354 for Street Gaussians, 0.322 for the variant labeled without pose optimization, and 0.253 for the full pipeline. The prose describes the full configuration as incorporating registration, bundle adjustment and pose optimization.

Go to primary source ↓
e-limitationsPDF p. 9, Limitations and Future Work, all six named subsectionsInspect

The authors do not establish correlation with real-world deployment. Per-scene reconstruction takes about 1–2 hours on unspecified current hardware. Remaining limitations include rendering artifacts, centerline-constrained rule-based traffic, unsafe human-expert exceptions, and empirically chosen weighting/aggregation.

Go to primary source ↓
e-trainingPDF p. 6, Section 4.1, Benchmark; p. 8, Section 4.3, Benchmark and Table 3 captionInspect

The correlation study evaluates existing rule-based and learned planner variants. The rendering diagnostic uses a pretrained LTF planner trained only on real data; no new joint planner-renderer training objective is specified.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.