PaperlayerView exact revision on arXiv

Source-linked AI summary

ExecDiffuser: Closing the Sample-to-Execution Gap in Diffusion Planning

Shuo Pei, Bosun Liang, Saihui Zha, Chen Sun, Huachun Tan, Yong Wang

arXiv:2610.36837v1cs.RO

TL;DR

Closed-loop driving depends on the trajectory committed to control, yet generative planners are often assessed by what they can generate, and stronger proposals alone can worsen the executed decision. ExecDiffuser couples execution-conditioned proposal learning with recurrent candidate selection and returns the published trajectory as next-cycle context. It reports strong closed-loop results across nuPlan, InterPlan, and NAVSIM, with clear gains under reactive and interaction-heavy conditions.

  • Problem

    Generative planners are commonly assessed by candidate quality, although closed-loop driving depends on the trajectory committed to control and proposal improvements alone can lower committed quality.

    Source: §I · p. 1

  • Method

    ExecDiffuser couples execution-conditioned hypothesis responsibility and trajectory shaping with recurrent, candidate-conditioned publication utility.

    Source: §I · pp. 1–2

  • Results

    ExecDiffuser achieves strong closed-loop results across nuPlan, InterPlan, and NAVSIM, with particularly clear gains under reactive and interaction-heavy conditions.

    Source: §V · p. 7

  • Takeaways & Limitations

    The findings suggest evaluating generative planners by both what they can generate and what they ultimately commit to control.

    Source: §V · p. 7

    Future work should address forecast uncertainty, false-valid decisions in occluded scenes, and validation in real-world interactive environments.

    Source: §V · p. 7

Abstract

from arXiv · show

Generative planners are often evaluated by the trajectories they can generate, although closed-loop driving ultimately depends on the single candidate committed to control. We identify this mismatch as the sample-to-execution gap and show that improving proposal quality alone can worsen the executed decision. Here, we introduce ExecDiffuser, a recurrent proposal-to-execution framework that couples candidate generation, trajectory selection, and subsequent replanning through shared execution semantics. Intent-Guided Diffusion Proposal (IGD) assigns hypothesis responsibility and shapes trajectories with execution-conditioned supervision. Execution-Grounded Selection (EGS) learns a candidate-conditioned publication utility, commits one trajectory to control, and returns it as context for the next planning cycle. Across nuPlan, InterPlan, and NAVSIM, ExecDiffuser achieves leading closed-loop performance against strong state-of-the-art baselines. Under a unified five-seed protocol, it obtains an average score of 85.62, improving over Diffusion Planner and FlowDrive by 4.02 and 1.31 points, respectively. On NAVSIM, it achieves 92.6 PDMS under the official evaluation pipeline. Matched counterfactual analysis further shows that ExecDiffuser more reliably converts strong candidates into committed decisions, while prefix interventions confirm that this advantage persists after physical execution and subsequent replanning.

I. INTRODUCTION

Closed-loop planning depends on the trajectory committed to control, not only on the quality of the generated candidate set. ExecDiffuser connects proposal learning, publication, and replanning through shared execution semantics.

II. RELATED WORK

  • Prior generative-planning advances improve proposal quality and coverage, but only one trajectory is ultimately committed to control.The related-work discussion also notes that existing execution-aware evaluation does not isolate candidate availability from the decision that publishes a candidate.Source: §I · p. 1, §II · p. 2

I. INTRODUCTION

  • A stronger proposal set can raise the best available candidate’s quality while lowering committed quality and increasing candidate-publication regret.With the selector fixed, oracle-best quality rises from 90.33 to 91.36, while committed quality falls from 89.82 to 89.64 and regret rises from 0.51 to 1.72.Source: §I · p. 1
  • Proposal responsibility and controller-facing selection can use mismatched evidence, although the committed trajectory changes the state from which replanning occurs.Proposal supervision may follow imitation proximity or geometric matching, while selection must use current execution evidence such as driveability, route adherence, interactions, progress, comfort, and history.Source: §I · p. 1
  • ExecDiffuser couples candidate generation, publication, and next-cycle replanning through shared execution semantics.IGD aligns hypothesis responsibility and trajectory shaping with execution-conditioned supervision; recurrent EGS learns publication utility and returns the selected trajectory as context.Source: §I · pp. 1–2
  • A matched counterfactual protocol separates available-best quality from published quality, while commit-horizon controls and resumed-replanning interventions test effects after execution.The paper reports a unified five-seed comparison against reproduced Diffusion Planner and FlowDrive baselines.Source: §I · pp. 1–2

II. RELATED WORK

  • ExecDiffuser connects proposal responsibility and final candidate scoring through shared execution semantics, alongside a matched counterfactual evaluator.This addresses the stated lack of alignment between separate assignment and scoring objectives.Source: §II · p. 2

A. Problem Formulation and Overview

At each planning cycle, the planner uses scene and route information together with the previous publication to generate and publish a trajectory. The formulation distinguishes candidate quality available before selection from the trajectory actually committed to control.

A. Problem Formulation and Overview

  • At cycle t, the planner receives scene, map, route, and previous-publication context, generates candidates, and EGS publishes one trajectory.The prior publication is returned as context for the next planning cycle.Source: A. Problem Formulation and Overview · p. 2
  • The sample-to-execution gap compares the best retrospective candidate with the trajectory published to control.Candidate-publication regret is their difference; the retrospective oracle includes candidates rejected by the online mask, so it captures false rejection and ranking errors.Source: A. Problem Formulation and Overview · p. 2
  • IGD improves candidate availability, while EGS ranks realized candidates and returns its publication as next-cycle context.The shared interface uses a common encoder and route context, and supplies predicted risk evidence.Source: A. Problem Formulation and Overview · pp. 2–3
  • The online interface uses current-cycle information and evaluates candidates with drive, route, lane, risk, progress, kinematic, comfort, and temporal evidence.Lower evidence values are preferred; temporal evidence is gated by whether the aligned previous publication remains suitable.Source: B. Shared Online Execution Interface · p. 3
  • The hard mask checks drivability and collision, while route progress, comfort, graded risk, and temporal consistency remain soft preferences.The collision check uses uncertainty-inflated predicted agent footprints and calibrated overlap and time-to-collision thresholds over a fixed look-ahead horizon.Source: B. Shared Online Execution Interface · p. 3
  • The rule-cost baseline minimizes a fixed aggregation of online evidence, whereas EGS learns utility from candidate, scene, evidence, and recurrent context.No future simulator state or candidate-specific reactive rollout is available online.Source: B. Shared Online Execution Interface · p. 3

C. Intent-Guided Diffusion Proposal

Intent-Guided Diffusion Proposal (IGD) gives active query–noise hypotheses route and recurrent context, then decodes them into physical trajectories. Its training separates responsibility assignment from trajectory shaping.

C. Intent-Guided Diffusion Proposal

  • Each active query receives route and recurrent context, one noise sample, and decoding into a physical trajectory.The decoder predicts normalized clean states, which a fixed differentiable operator reconstructs as trajectory states.Source: C. Intent-Guided Diffusion Proposal · p. 3
  • IGD separates hypothesis ownership from trajectory correction, deriving training preferences from online execution evidence and training-only imitation proximity.The training preference includes imitation proximity, which is unavailable at inference.Source: C. Intent-Guided Diffusion Proposal · p. 3
  • The responsibility loss assigns a query–noise hypothesis, while a separate shaping loss updates its physical trajectory.Differentiable surrogates shape route, lane, soft-overlap, progress, kinematic, comfort, and temporal evidence; discrete collision and off-road indicators remain detached.Source: C. Intent-Guided Diffusion Proposal · p. 3
  • A scene-conditioned allocation head selects 2, 4, or 8 queries before decoding the highest-priority ones.Specialization regularizes query usage.Source: C. Intent-Guided Diffusion Proposal · p. 3

D. Recurrent Execution-Grounded Selection

EGS scores admitted candidates using online-observable scene, route, forecast, execution, and recurrent context, then publishes one trajectory for use in the next cycle. Experiments separately assess candidate selection and proposal quality under matched evaluation conditions.

D. Recurrent Execution-Grounded Selection

  • EGS combines each detached candidate trajectory with scene, previous-publication, execution-evidence, and continuity features to score publication quality.Its training preference supervises relative publication quality but is not an inference-time energy or a source of coordinate gradients through ranking.Source: D. Recurrent Execution-Grounded… · p. 3
  • Unlike a fixed linear rule-cost aggregation, EGS infers context-dependent utility from online-observable candidate geometry, scene tokens, route and forecast evidence, and the previous publication.No future simulator state or candidate-specific reactive rollout is available online.Source: D. Recurrent Execution-Grounded… · pp. 3–4
  • The selected trajectory is stored and returned to IGD and EGS at the next cycle; if every candidate is invalid, the planner invokes a deterministic jerk-limited MRM.Gradients do not pass across planning cycles.Source: D. Recurrent Execution-Grounded… · p. 4, Figure 2 · p. 4

IV. EXPERIMENTS

  • The experiments ask whether the planner improves closed-loop performance, whether IGD improves candidate availability, whether EGS improves publication, and whether initial commitment affects replanning.The candidate-level diagnosis executes candidates independently for an 8.0-second horizon; selected and oracle-best candidates share the initial simulator state and matched rollout conditions.Source: §IV · p. 4, A. Experimental Setup · p. 4

A. Experimental Setup

  • The reproduced baselines and ExecDiffuser share nuPlan versions, scenario lists, simulator settings, metric implementation, and five training seeds.The protocol reports arithmetic means without seed selection or ensembling, with confidence intervals from 10,000 paired scenario-level bootstrap resamples.Source: A. Experimental Setup · p. 4

B. Closed-Loop Results on nuPlan and InterPlan

ExecDiffuser achieves leading closed-loop results on nuPlan and InterPlan. Its ablations show that stronger proposals alone can worsen publication, while execution-aware selection and recurrent context improve candidate commitment and subsequent replanning.

B. Closed-Loop Results on nuPlan and InterPlan

  • 85.62 average score across six nuPlan settings leads the reproduced methods, exceeding Diffusion Planner by 4.02 points and FlowDrive by 1.31.The gains are largest in reactive settings, where committed trajectories change subsequent interactions.Source: B. Closed-Loop Results on nuPlan and… · p. 4, Table I · p. 5
  • 38.35 on interaction-intensive InterPlan improves over Diffusion Planner by 55.2% and FlowDrive by 3.8%.Together with the nuPlan gains, this extends the result beyond one simulator mode.Source: B. Closed-Loop Results on nuPlan and… · p. 5

C. Ablation and Closed-Loop Diagnosis

  • Oracle-best quality rises from 90.33 to 91.36, but published quality falls from 89.82 to 89.64 under the same imitation selector.Candidate-publication regret consequently expands from 0.51 to 1.72; the direction holds for every training seed.Source: C. Ablation and Closed-Loop Diagnosis · p. 5, Figure 3 · p. 5
  • With EGS fixed, IGD raises oracle-best quality from 91.36 to 92.63, increases feasible candidates from 86.60% to 92.10%, and reduces all-invalid sets from 2.17% to 0.82%.Specialization further raises the ceiling to 92.93 and lowers all-invalid sets to 0.68%; the gain persists at a matched average decoded budget of 5.5.Source: C. Ablation and Closed-Loop Diagnosis · p. 5, Table II · p. 6
  • Full EGS reaches 86.6% oracle top-1 recall and 0.62 regret, a 30.2-point recall gain and 75.2% regret reduction over imitation.With proposals fixed, the cascade progressively improves recall and reduces regret as scoring, online evidence, hard validity, and recurrent context are added.Source: C. Ablation and Closed-Loop Diagnosis · p. 5, Figure 4 · p. 5, Table II · p. 6
  • After different initial publications, the same resumed planner gains 0.51 to 1.90 points over imitation as the prefix lengthens, with every paired interval excluding zero.At a 4.0-s prefix, the EGS advantage ranges from 0.39 points in straight following to 2.04 points in unprotected turns.Source: C. Ablation and Closed-Loop Diagnosis · pp. 5–6, Figure 5 · p. 6

D. Cross-Benchmark Evaluation on NAVSIM

ExecDiffuser leads the compared methods on several NAVSIM metrics, and its reliability and conclusion results highlight both improved execution and a remaining forecast-dependent failure boundary. The paper concludes that proposal quality and committed-decision quality are distinct, and that initial commitments affect subsequent replanning.

D. Cross-Benchmark Evaluation on NAVSIM

E. Reliability, Failure Boundaries, and Cost

  • The false-valid rate rises from 2.70% overall to 5.14% in occluded intersections, marking a forecast-dependent failure boundary.The Adaptive Continuity Gate reduces plan switching from 13.6% to 7.3%, with 0.18 s hazard-response delay and 93.4% hazard success.Source: E. Reliability, Failure Boundaries,… · p. 6
  • ExecDiffuser adds 5.2M parameters and runs at 59.5 ms, versus 50.7 ms for Diffusion Planner and 47.9 ms for FlowDrive.The paper also reports fewer unnecessary stops, high-jerk events, and emergency braking than Diffusion Planner.Source: E. Reliability, Failure Boundaries,… · pp. 6–7

V. CONCLUSION

  • With the selector fixed, stronger proposal learning raises best-candidate quality but can lower committed-trajectory quality and enlarge publication regret.The conclusion identifies a higher proposal ceiling as no guarantee of a better closed-loop decision.Source: §V · p. 7
  • Controlled ablations attribute candidate availability to execution-aware proposal learning and candidate-to-decision conversion to execution-grounded selection.Prefix interventions show that different initial commitments lead to different physical states and subsequent replanning performance and recovery behavior.Source: §V · p. 7

APPENDIX A PROTOCOL AND REPRODUCIBILITY

The appendix describes fixed, shared evaluation settings and the implementation choices linking candidate generation, selection, and recurrent alignment. It also specifies a counterfactual evaluator, eligibility rules, and scenario-level uncertainty estimation.

A. Training and Evaluation Configuration

  • Calibration statistics, thresholds, and loss coefficients are fixed before benchmark runs, with identical scenario seeds, simulator settings, and metric implementations for each candidate and method pair.These settings are specified as shared across all seeds and controlled ablations.Source: A. Training and Evaluation Configuration · p. 8, Appendix A · p. 8

B. Candidate Representation, Recurrent Alignment, and Gradient Routing

  • Training preference and online evidence are evaluated on the same physical trajectory representation, while shaping updates proposal coordinates and ranking gradients do not pass through candidate coordinates.The previously published trajectory is aligned to the current ego frame, its executed prefix is removed, and its unexecuted suffix is resampled as recurrent state without gradients across cycles.Source: B. Candidate Representation, Recurrent… · p. 8

C. Counterfactual Evaluator and Diagnostic State Set

  • The counterfactual evaluator commits each candidate for 8.0 s from a cloned state, reinitializing reactive agents for each rollout and scoring selected and oracle-best candidates with the same metric aggregation.Its diagnostic measures publication regret, including ranking errors and false rejection by hard validity; it is excluded from training, calibration, checkpoint selection, and online inference.Source: C. Counterfactual Evaluator and… · pp. 8–9
  • Diagnostic states exclude incomplete candidate sets, masks, rollout outputs, or horizons; the five seeds contribute 112,840 eligible states in total.Method-comparison confidence intervals use 10,000 paired scenario-level bootstrap resamples, not planning states as independent units.Source: C. Counterfactual Evaluator and… · pp. 8–9

D. Execution Evidence, Normalization, and Hard Validity

The planner constructs normalized execution evidence with a shared penalty direction and applies a deliberately narrow hard-validity mask. Its continuity gate modulates temporal evidence without overriding that mask.

D. Execution Evidence, Normalization, and Hard Validity

  • Normalized execution evidence uses a common orientation: zero is favorable and larger values are worse.The evidence enters training preference, rule-cost control, and learned evidence encoding; coordinate shaping uses aligned differentiable surrogates.Source: D. Execution Evidence, Normalization,… · p. 9, Table A2 · p. 10
  • The hard-validity mask rejects candidates failing drivable-region or imminent-collision checks; route deviation, lane consistency, progress, comfort, and moderate risk remain soft evidence.Collision checks use uncertainty-inflated predicted agent footprints over a fixed look-ahead horizon.Source: D. Execution Evidence, Normalization,… · p. 9
  • Feasible APD is computed from hard-valid candidates, and states with fewer than two such candidates are excluded rather than assigned zero diversity.Its value is interpreted alongside coverage, feasible-candidate ratio, and the all-invalid rate.Source: D. Execution Evidence, Normalization,… · pp. 9–10

E. Adaptive Continuity Gate

  • The Adaptive Continuity Gate uses scene and aligned-prior-publication evidence to modulate temporal evidence, but never overrides hard validity.The training target uses the previous publication's hard-validity result and normalized risk cost.Source: E. Adaptive Continuity Gate · p. 10

F. Shared Shaping Weights and Query Specialization

The appendix separates responsibility assignment from coordinate shaping, then describes scene-conditioned query allocation and training-only forecast supervision. At inference, allocation chooses queries before decoding, while future ground-truth agent states are not used.

F. Shared Shaping Weights and Query Specialization

  • Execution-aligned shaping uses a coordinate-differentiable surrogate, while discrete collision and off-road indicators remain detached; query specialization supports coverage without replacing execution-conditioned responsibility.The global shaping coefficient controls this coordinate-level path.Source: F. Shared Shaping Weights and Query… · pp. 10–11

G. Scene-Conditioned Candidate Allocation

  • The priority head uses only ht, ct, and qm—not decoded candidates, candidate evidence, or EGS scores—and selects top-Kt queries before diffusion decoding, providing a real compute reduction.A smaller budget is selected only when the training-prioritized prefix preserves the best full-budget hard-valid cost within ϵK; at inference, the learned scene-conditioned order replaces the training-preference order.Source: G. Scene-Conditioned Candidate… · p. 11

H. Forecast Supervision

I. Fixed Verification, Training, Allocation, and MRM Parameters

The controlled evaluations separate candidate availability from publication quality and test whether publication differences persist across commit horizons and after execution-driven replanning. Results show that better proposals alone need not improve committed decisions, while EGS more reliably selects strong candidates.

J. Baseline Reproduction and Compute Cost

A. Five-Seed Consistency of the Proposal-Only Contradiction

  • Proposal-only IGD raises oracle-best quality for all five seeds but slightly lowers published quality and increases regret under the unchanged imitation selector.The available-best improvement is 1.03 points, while regret increases by 1.21 points and oracle top-1 recall falls by 6.8 percentage points.Source: A. Five-Seed Consistency of the… · p. 11, Table A7 · p. 13

APPENDIX B MAIN-TABLE RESULTS AND STATISTICAL RELIABILITY

APPENDIX C PROPOSAL AND PUBLICATION ABLATIONS

  • Controlled ablations separate proposal learning, publication interfaces, and recurrent continuity; figures visualize the proposal and publication cascades.Proposal-side rows hold EGS fixed, publication-side rows freeze the proposal generator, and temporal rows vary only continuity.Source: Appendix C · p. 13, Figure A2 · p. 13, Figure A3 · p. 14

A. Commit-Horizon Sensitivity with a Fixed Candidate Set

  • EGS achieves the highest published score and lowest regret at every commit horizon, reducing regret by 66.9%–75.2% versus imitation.Its top-1 recall remains 86.1%–88.1%; the publication advantage is not specific to the original 8.0-s horizon.Source: A. Commit-Horizon Sensitivity with a… · pp. 13, 15

B. Prefix Intervention with Resumed Replanning

  • Prefix interventions force a candidate for H seconds, then resume the same frozen planner from the physical and recurrent state induced by execution.Branches use matched cloned states and a fixed 8.0-s total evaluation window.Source: B. Prefix Intervention with Resumed… · p. 15
  • The retrospective oracle prefix is selected by total resumed-branch score over the full 8.0-second window and is unavailable to training and online inference.It is a diagnostic upper bound, not an online selection method.Source: B. Prefix Intervention with Resumed… · p. 15, Table A13 · p. 16
  • EGS-minus-imitation resumed-score gains increase from 0.51 to 1.90 points as forced-prefix horizons grow from 0.5 to 4.0 seconds.TTC differences likewise rise from 0.60 to 2.27 points; all paired resumed-score intervals exclude zero.Source: B. Prefix Intervention with Resumed… · p. 16, Table A13 · p. 16

C. Scenario-Wise Effects of the Initial Prefix

Prefix interventions show that EGS's downstream advantage is concentrated in interaction-heavy scenes. The appendix further examines recurrent stability, forecast-dependent fallback limits, computational cost, and candidate-allocation controls.

C. Scenario-Wise Effects of the Initial Prefix

A. Does Recurrent Publication Remain Stable Under Forecast Failure?

B. Where Do the Gains Occur, and What Do They Cost?

D. Matched-Compute Budget and Candidate Priority

G. MRM Benefit and Forecast-Failure Coverage

H. High-Risk False-Valid Analysis

The appendix reports execution diagnostics, risk-margin trade-offs, scenario-dependent gains, and NAVSIM evidence. NAVSIM results are supported by five independent runs and qualitative logged-state replanning examples, not closed-loop rollouts of the displayed trajectories.

I. Official Val14 Sub-Metrics

  • Across five independently trained models, ExecDiffuser obtains an official Val14 score of 92.37, with TTC 94.06 and route progress 90.05.The reported sub-metrics also include collision 100.00, comfort 90.99, drivable-area compliance 100.00, driving-direction compliance 99.57, and speed-limit compliance 100.00.Source: §I · p. 20

J. Non-Conservative Execution

  • The full model raises TTC from 86.14 to 94.06 and route progress from 87.80 to 90.05 while maintaining average speed near Base Diffusion.Unnecessary stops, high-jerk events, and emergency braking fall from 4.82/5.41/2.16% to 3.19/3.59/1.12%.Source: J. Non-Conservative Execution · p. 20, Figure A9 · p. 21

K. Risk-Margin Sensitivity

  • The default risk margin of 0.50 jointly maximizes NR, R, and comfort; larger margins improve TTC but reduce route progress and overall closed-loop performance.The reported sensitivity identifies a safety–efficiency trade-off rather than a monotonic tuning gain.Source: K. Risk-Margin Sensitivity · p. 20, Figure A9 · p. 21

L. Scenario-Wise Performance

  • Gains over FlowDrive are largest in unprotected turns, pedestrian or cut-in events, dense traffic, and merge or yield scenarios, and smallest in straight following.The appendix frames this concentration as consistent with execution-grounded publication mattering most when interaction changes trajectory value.Source: L. Scenario-Wise Performance · p. 20, Figure A9 · p. 21

A. Protocol and Seed-Level Results

B. Temporal Qualitative Analysis

  • The five-frame NAVSIM excerpts show selected trajectories across consecutive logged observations, with 0.5-s spacing over 2.0 s.Orange denotes the logged expert and cyan the selected trajectory; the sequences illustrate replanning consistency, not closed-loop execution.Source: B. Temporal Qualitative Analysis · p. 21, Figure A11 · p. 22
  • The NAVSIM excerpts do not establish a causal effect of a published trajectory on the next logged state because each prediction uses its corresponding logged observation.The evidence is qualitative local consistency, while quantitative claims rely on official navtest evaluation and per-run results.Source: B. Temporal Qualitative Analysis · pp. 21–22

Related papers

Loading 2610.36837v1…