acceptodds
Under review as a conference paper at ICLR 2027

F2-Planner: Generative Replanning from Unexecuted Future Trajectories

Abstract

Generative trajectory planners repeatedly refine future trajectories, although only a short prefix of each selected plan is executed before replanning. The unexecuted future offers a structured starting point, but vehicle motion and new observations can make it outdated. We propose F2-Planner, a future-to-future generative replanner that carries the unexecuted portion of the controller-selected trajectory across cycles as a revisable generative state. It first advances the previous plan and aligns the remaining waypoints with the current ego pose. A horizon-wise utility predictor then controls how much of this future is blended with current candidates and how broadly the resulting source is sampled. On Bench2Drive v0.0.4, F2-Planner achieves a mean Driving Score of 96.65 with two neural function evaluations across eight training seeds. It outperforms the matched four-evaluation cycle-local planner (96.01) while reducing planning-head latency by 27.77%. At the same two-evaluation budget and nearly matched planning-head latency, it improves over an execution-aligned warm-start by 0.42 Driving Score points. Separate event-response and corrupted-history tests show that the carried future can be revised when its intent becomes stale. The source update also improves a truncated-diffusion generator within the same planning stack. These results support reusing and revising an unexecuted future to reduce repeated refinement while retaining responsiveness in the evaluated scenarios.

Primary area:
applications to robotics, autonomy, planning
Keywords:
autonomous driving, generative planning, diffusion models, trajectory planning, recurrent planning, closed-loop planning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dGjBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes F2-Planner, which reduces redundant computation in autonomous driving by reusing the unexecuted portion of the previous trajectory. It first aligns the previous trajectory with the vehicle’s actual motion, then predicts waypoint-wise inheritance weights to jointly control its blending with current candidates and the sampling noise. A few generative refinement steps produce the new trajectory.On Bench2Drive, the method outperforms a four-evaluation baseline with only two neural function evaluations, reducing planning-head latency by 27.77%.

Critical strengths
  1. The problem is practically relevant. The paper addresses redundant generation costs in receding-horizon planning by leveraging the unexecuted portion of the previous trajectory to improve initialization, offering a reasonable approach to reducing inference costs in generative planning.
  2. The method is clearly designed and delivers measurable gains. Waypoint-wise inheritance weights jointly control the source distribution’s center and sampling spread, preserving existing driving intent while allowing outdated plans to be revised. With only two neural function evaluations, the method achieves a higher Driving Score than the four-evaluation baseline while reducing planning-head latency.
Critical weaknesses
  1. Response and recovery capabilities under a low computational budget are insufficiently evaluated. The paper’s main efficiency gains are demonstrated with two neural function evaluations, whereas event-response and recovery experiments involving erroneous history use four evaluations. It therefore remains unclear whether the highlighted low-latency configuration can promptly revise outdated or incorrect plans.
  2. Generalizability and system-level benefits remain unclear. The experiments use the same perception and control stack, and the results with a second generator do not establish transfer across different system architectures. Moreover, latency measurements cover only the planning head and are therefore insufficient to quantify the actual speedup of the complete driving system.
Questions for authors
  1. How does reducing inference from four to two NFEs affect event response and recovery from erroneous history? Do you have results for the two-NFE configuration showing whether the reported efficiency gains preserve these capabilities?
  2. What fraction of the complete sensor-to-control latency does the planning head account for, and what end-to-end latency reduction does the two-NFE configuration achieve?
3 comments
AuthorsOpenReview ↗

Response to Reviewer dGjB

Thank you for recognizing the practical motivation, methodological design, and computational benefits of our work. Following your feedback, we evaluated response and recovery at 2 NFEs, measured system-level latency, and validated our method on an independent planning system (references denote the revised manuscript).

1. Response and recovery when reducing from 4 to 2 NFEs.

We repeated the event-response and corrupted-history recovery experiments at 2 NFEs using the original event definitions, criteria, fault injections, and evaluation windows. Each configuration evolves using its own closed-loop history under matched event identities and exogenous conditions, with exactly two generator evaluations per cycle and no extra calls for F2 (Appendix K.1, p. 34). Compared against same-budget aligned warm-start (AWS), F2 reduces longitudinal and abrupt-intent response RMST by 39 ms and 25 ms (Tables A35–A36, pp. 34–35), and reduces single-fault and sustained-drift recovery RMST by 0.95 s and 1.14 s (Tables A37–A38, p. 35). All four paired 95% CIs lie strictly below zero, with RMST accounting for all non-responding and unrecovered cases.

Reducing F2 from 4 to 2 NFEs modestly increases RMST point estimates—for example, from 0.464 to 0.493 s for abrupt-intent response (Tables A27, A35, pp. 29, 34) and from 1.46 to 1.75 s for sustained-drift recovery (Tables A31, A37, pp. 33, 35). Because these cross-budget comparisons are descriptive (paired CIs for F2-2 vs. F2-4 are not reported), we do not claim that halving the budget is behaviorally cost-free or statistically equivalent to 4 NFEs. Rather, these results directly verify that F2 maintains effective response and recovery at 2 NFEs while achieving consistently lower RMST than same-budget AWS.

2. Planning-head contribution and end-to-end latency benefits.

We profiled the complete observation-to-control-command chain—covering input transfer and preparation, perception, planning, control computation and submission, and waiting time—excluding physical actuator dynamics (Appendix Q.1, p. 40). In Cycle-local-4, the planning head takes 47.31 ms, accounting for 40.18% of its 117.74 ms mean total latency (Table A50, p. 41). F2-2 lowers mean total latency from 117.74 to 104.10 ms, yielding an end-to-end reduction of 13.64 ms (11.58%; Tables A50, A52, p. 41). Of this reduction, 13.14 ms comes directly from the planning head, while the remaining 0.50 ms reflects minor non-planning runtime variance rather than algorithmic acceleration of perception or control.

Because perception and other non-planning stages remain fixed in the chain, the 11.58% end-to-end reduction is naturally smaller than the 27.77% relative reduction within the planning head. Within F2, moving from 4 to 2 NFEs lowers mean total latency from 120.95 to 104.10 ms (13.93%). At the matched 2-NFE budget, F2-2 takes 0.75 ms longer than AWS-2 due to the source constructor, so we do not claim a same-budget speed advantage (Tables A50–A52, p. 41). Table A51 (p. 41) reports P50, P95, and P99 latencies as descriptive runtime measurements under our test setup, not confidence intervals or cross-hardware guarantees.

3. Applicability beyond the original planning stack.

To evaluate cross-system applicability beyond replacing a generator within the same stack, we integrated F2 into Diffusion Planner on the nuPlan closed-loop benchmark, retaining its native scene encoding, joint prediction architecture, and control interface (2 NFEs, 8 matched training seeds, identical training data, budgets, post-processing, and controllers; Appendix L.1, pp. 35–36). Across Test14-random and Test14-hard under both non-reactive and reactive traffic, F2 improves over AWS by 0.9–1.5 points on the 100-point scale, with all four pointwise paired 95% CIs strictly above zero (Tables A39–A40, p. 36). Using a shared adaptation interface isolates the adaptive source update from aligned history reuse. These results extend our empirical validation to a distinct planning system and benchmark, supporting application after adaptation and matched source–refiner training rather than training-free transfer to arbitrary planners.

Thank you again for prompting us to connect planning-head efficiency with measured system-level latency and to directly evaluate response and recovery under the low-budget regime. We hope these additions thoroughly address your concerns.

We would like to highlight two findings from the revision that directly address your main concerns: F2 can revise stale history at the two-NFE budget, and its computational savings remain measurable beyond the planning head.

At two NFEs, F2 responds and recovers faster than same-budget aligned warm-start. Each configuration uses its own closed-loop history under matched event and fault settings. F2 reduces response RMST by 39 ms and 25 ms in the longitudinal and abrupt-intent suites, and recovery RMST by 0.95 s and 1.14 s under single faults and sustained drift. All four paired 95% confidence intervals lie below zero (Tables A35–A38). RMST includes nonresponding and unrecovered cases, so these improvements are not restricted to successful cases.

Some RMST estimates increase when reducing F2 from four to two NFEs. Nevertheless, the direct two-NFE comparisons show that F2 retains the ability to revise outdated plans and improves on conventional aligned history reuse under the same refinement budget. This supplies the responsiveness evidence that was missing from our original quality–compute comparison.

The planning-head savings translate into an 11.58% reduction in measured observation-to-control-command latency. The baseline planning head accounts for 40.18% of total mean latency. After including input preparation, perception, control computation, command submission, and waiting time, F2-2 reduces the mean from 117.74 ms to 104.10 ms relative to Cycle-local-4. Its measured P95 and P99 latencies are also lower (Tables A50–A52). Thus, the broader processing costs reduce the relative speedup, but a system-level benefit remains under the measured setup.

The added Diffusion Planner/nuPlan evaluation also extends the evidence beyond the original stack. With the adaptation interface and two-NFE budget held fixed, F2 improves over aligned warm-start by 0.9–1.5 score points across all four settings, with each paired confidence interval above zero (Tables A39–A40). This supports applicability after adaptation and matched training.

We hope these results address the main concerns behind your assessment, and would appreciate your reconsideration of the rating in light of the revised evidence. Thank you for reviewing these additions.

Thank you for asking us to evaluate response and recovery at the same budget used for our efficiency results. Reducing F2 from four to two generator evaluations per cycle cuts mean system latency by 13.93% and increases the response and recovery time estimates. At the same two-NFE budget, F2 still achieves lower response and recovery RMST than aligned warm-start in all four tests.

Each configuration uses its own closed-loop history. RMST averages the time to respond or recover across all cases, counting cases with no response or recovery at the evaluation window limit. AWS denotes aligned warm-start.

Metric ↓F2-4F2-2AWS-2F2-2 − F2-4F2-2 − AWS-2 (paired 95% CI where available)
Longitudinal response RMST0.463 s0.473 s0.513 s+10 ms−39 ms [−58, −17]
Abrupt-intent response RMST0.464 s0.493 s0.518 s+29 ms−25 ms [−42, −5]
Single-fault recovery RMST1.54 s1.84 s2.79 s+0.30 s−0.95 s [−1.33, −0.55]
Sustained-drift recovery RMST1.46 s1.75 s2.89 s+0.29 s−1.14 s [−1.54, −0.73]
Mean observation-to-control-command latency120.95 ms104.10 ms103.35 ms−16.85 ms (−13.93%)+0.75 ms

All four paired RMST intervals for F2-2 versus AWS-2 lie below zero. The comparisons across NFE budgets and all latency differences are descriptive. The −39 ms longitudinal effect uses unrounded values (Tables A27, A31, A35–A38, and A50–A52).

Your question about the planning head’s share of total latency also prompted us to measure the full chain from observation to control-command submission. The planning head accounts for 40.18% of mean latency in Cycle-local-4. Compared with that baseline, F2-2 reduces the mean from 117.74 to 104.10 ms (11.58%), while increasing the Bench2Drive Driving Score from 96.01 to 96.65.

To address dependence on the original stack, we also evaluated F2 in Diffusion Planner on nuPlan. After adaptation and matched training, F2 improves over the same-budget aligned warm-start by 0.9–1.5 score points across both test splits and both traffic modes, with the adaptation interface held fixed (Tables A39–A40).

We hope these comparisons answer your question about what is gained and what is lost at two NFEs. Thank you for prompting us to measure that trade-off directly.

Reviewer_H5YwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes F2-Planner, a future-to-future generative replanning framework for autonomous driving. The central observation is that, in a receding-horizon planning system, only a short prefix of the selected trajectory is executed before the next planning cycle, while the remaining future contains useful information about the current driving intent. Rather than regenerating each trajectory from a cycle-local source, F2-Planner carries the unexecuted portion of the previously selected trajectory into the next cycle.

The method first advances the previous controller-selected trajectory according to estimated path progress and transforms the remaining waypoints into the current ego frame. It then predicts a waypoint-wise inheritance vector that determines how much of the reconciled previous future should contribute to the next generative source. This inheritance value simultaneously controls the source mean and sampling spread: high inheritance preserves and concentrates around the previous future, while low inheritance shifts the source toward current-cycle candidates and increases exploration. The resulting source is refined using a generative trajectory planner.

Critical strengths
  • The paper includes a capacity-matched history-interface control and reports similar performance between execution-aligned conditioning and aligned warm-start. This is useful because it prevents the conclusion that any access to previous history automatically provides the reported gain. The component ablation also suggests that the combination of center adaptation and adaptive spread is more effective than either component alone under the reported low-NFE settings.

  • The event-response and recovery experiments are valuable. They address a central concern with recurrent planning: a cached future can become stale or harmful after a cut-in, braking event, or intent change. The paper reports that learned inheritance decreases in stale-intent events and that the source spread increases, which is consistent with the intended behavior.

  • The paper identifies a real inefficiency in receding-horizon generative planning. A full future trajectory is generated at every cycle, although only a short prefix is executed. Reusing the unexecuted future is a natural way to reduce repeated computation, and the paper frames this idea clearly. This is more meaningful than simply caching an internal latent state because the carried trajectory is the actual controller-selected plan. It therefore reflects the planner’s recent decision and not merely an intermediate representation.

  • The paper includes a capacity-matched history-interface control and reports similar performance between execution-aligned conditioning and aligned warm-start. This is useful because it prevents the conclusion that any access to previous history automatically provides the reported gain. The component ablation also suggests that the combination of center adaptation and adaptive spread is more effective than either component alone under the reported low-NFE settings.

Critical weaknesses
  1. The core method combines several familiar components: (1) align a previous trajectory with the current ego frame; (2) estimate progress along the previous path; (3) blend the aligned trajectory with current candidate trajectories; (4) adjust sampling variance based on the blend weight; (5)refine the source using a generative planner. But the paper should clearly establish what is fundamentally new beyond existing warm-start, temporal diffusion, trajectory-anchor, and diffusion-bridge methods.

  2. The paper explicitly reports a source–refiner crossover, indicating that the recurrent source is not universally beneficial when transferred to a different refiner or training distribution. This is an important result, but it weakens the broader interpretation of F2-Planner as a generally reusable replanning principle. The main claims should therefore be scoped more carefully. The evidence supports: F2-Planner improves the quality-compute frontier when the source constructor and generative refiner are trained as a matched pair within the evaluated planning stack.

An execution-reconciled future source is generally superior across generative planners or can be inserted as a plug-and-play module. The transfer experiment with the truncated-diffusion generator is encouraging, but it appears to use the same shared planning stack and does not establish robustness across substantially different architectures, datasets, candidate representations, or controllers. A stronger evaluation would apply the source mechanism to at least one independent generative planner with substantially different parameterization and training procedure, without extensive retuning.

  1. The main evaluation uses Bench2Drive v0.0.4 and CARLA 0.9.15. This is a relevant benchmark, but its 220 routes and 44 scenario types represent a limited closed-loop environment compared with the diversity of real-world driving. The claims about efficient generative replanning and stale-intent handling would be more convincing with at least one additional closed-loop benchmark or simulator, such as: nuPlan closed-loop evaluation, Waymax or WOMD-based simulation, a different CARLA based dataset like Fail2Drive, Longest6. Without such tests, it is difficult to determine whether the improvement reflects the general source construction or adaptation to the specific Bench2Drive dynamics and evaluator.

  2. The inheritance vector controls both source location and exploration scale. Therefore, a high utility value can cause the generator to remain concentrated around a previous future even when that future is unsafe or behaviorally stale. Conversely, a low utility value broadens sampling but may increase trajectory variability. The paper explicitly states that utility is a geometric inheritance signal and not a safety estimate. This is a critical limitation for deployment. The manuscript should more directly analyze:

collision rate under stale-history perturbations; off-road and lane-departure rates; TTC and minimum-distance distributions; failure cases where high inheritance persists despite a newly emerging hazard; failure cases where excessive exploration destabilizes the planner. The qualitative example with a cut-in is not sufficient to establish safety. The paper should report safety metrics specifically for the event-response and corrupted-history subsets.

  1. The two-NFE F2-Planner achieves a higher DS than Cycle-local-4, but the Success Rate difference is accompanied by a confidence interval that includes zero:

SR difference: 1.53 percentage points, 95% CI [−1.153, 4.221].

This means the strongest statistical evidence is for the Driving Score and latency improvement, not necessarily for success rate. The abstract and conclusion should avoid implying a broad improvement in all closed-loop outcomes.

Similarly, the four-NFE response and recovery experiments report lower restricted mean response/recovery times, but the recovery-rate intervals include zero. The authors correctly mention this, but the distinction should be more prominent in the main claims.

  1. The method carries a generated trajectory between planning cycles. Because Bench2Drive routes are temporally correlated, the effective number of independent samples is much smaller than the number of planning cycles. The paper reports eight training seeds and 220 routes per seed, which is useful, but it should clarify:

whether route-level results are paired across methods; whether all methods use exactly the same simulator seeds and traffic-agent seeds; how route-level aggregation is performed; whether training seeds are independent from simulator seeds; whether any route or scenario information is implicitly reused through the cached state; whether the cache is reset at route boundaries and scenario restarts. The manuscript mentions matched seeds and paired intervals, but reproducibility would improve with a precise hierarchy of replication units and aggregation rules in the main paper.

Questions for authors

The notation around publication-time trajectory, current trajectory, and transformed trajectory is dense. A compact algorithm box would make the procedure easier to reproduce.

The method uses “utility” for ρ t ρ t ​ , but the quantity is not a utility or value estimate in the decision-theoretic sense. “Inheritance weight” or “reuse confidence” would be more precise.

The source center is computed independently at each waypoint. The authors state that this need not be smooth before refinement, but it would be useful to quantify the magnitude of source discontinuities and confirm that the refiner consistently removes them.

The target in Equation 8 is based on the highest-scoring current candidate. The paper should explain how ties, invalid waypoints, and missing candidate support are handled.

The paper should report whether the source constructor adds meaningful parameter count or memory overhead relative to the baseline.

The transfer experiment would be easier to interpret if the exact degree of retuning for the truncated-diffusion generator were stated.

The paper reports published Bench2Drive context rows together with controlled rows. These should be visually separated to avoid suggesting that all rows were reproduced under exactly the same software, data, and evaluation conditions.

The qualitative examples are useful but should include failure examples where F2-Planner retains stale intent or overreacts to noisy observations.

The paper refers to a “controller-selected trajectory,” but the relationship between the planner selector and the low-level controller should be described more explicitly. It matters whether the controller tracks the selected trajectory exactly or introduces systematic deviations.

The paper should clarify whether the inherited trajectory is generated by the same model being evaluated or by a separately trained previous checkpoint.

3 comments
AuthorsOpenReview ↗

Response to Reviewer H5Yw

Thank you for recognizing our unexecuted-future motivation, history controls, and event-response experiments. Your feedback helped clarify our scope and motivated the analyses below (references denote the revision).

1. Contribution beyond conditioning and warm-start.

Alignment, history reuse, and refinement are not individually new. Rather, F2 builds a scene-adaptive recurrent source from the execution-reconciled unexecuted controller-facing trajectory, with learned waypoint-wise inheritance jointly controlling center and spread. Capacity-matched aligned conditioning and aligned warm-start (AWS) score similarly at 4 NFEs (96.69 vs. 96.71 DS; Table A3), so no generic source-interface advantage over conditioning is claimed. Against same-budget AWS and component controls at 2 NFEs, F2 gains 0.42 DS over AWS (paired 95% CI [0.130, 0.710]; Table A16); center/spread ablations support the joint update (Table 2).

2. Independent system and scope.

We evaluated Diffusion Planner on nuPlan, retaining its scene encoding, joint prediction architecture, and control interface (2 NFEs, 8 matched seeds, identical data, budgets, post-processing, and controllers; Appendix L.1). Across Test14-random/hard and non-reactive/reactive traffic, F2 gains 0.9–1.5 points over AWS on the 100-point scale (each pointwise paired 95% CI > 0; Tables A39–A40). Fixing the shared adaptation interface between F2 and AWS isolates the adaptive source update. With the source–refiner crossover (Tables A21–A22), this supports benefits after adaptation and matched source–refiner training, not universal plug-and-play transfer. In truncated diffusion (Appendix J), only the pre-noising source changes while the shared stack, noise schedule, and denoising updates stay fixed—a controlled generator substitution, not training-free transfer.

3. Safety and failure diagnostics.

On the 2-NFE event-response and corrupted-history subsets, Table A42 reports collision, off-road, and lane-departure counts/denominators, including non-responding and unrecovered cases: collisions drop from 19/800 to 12/800 (abrupt intent) and 9/160 to 6/160 (sustained drift) from AWS to F2. Table A43 gives median case-minimum TTC, body-to-body distance, and valid-TTC coverage; Tables A44–A45 give paired TTC summaries on common valid cases (median differences +0.06 to +0.15 s) and invalid-TTC reasons. These summaries do not provide full distributions or establish statistical safety gain/non-inferiority. Tables A47–A49 examine persistent high inheritance and low inheritance with instability: under sustained drift, high inheritance occurs in 11/160 cases (3 unrecovered, 3 collisions). These associations neither prove causation nor replace individual failure visualizations.

4. Trajectory geometry.

Table A46 measures spatial geometry of refined and selected trajectories on matched frozen states. For selected outputs, F2 vs. AWS has mean seed-level P95 of 5.3° vs. 5.2° (absolute turns) and 0.028 vs. 0.027 m⁻¹ (adjacent-curvature differences). F2's values are slightly higher; as they measure output geometry rather than paired pre/post-refinement source discontinuities, we do not claim universal removal of irregularities or statistical smoothness equivalence.

5. Statistical claims and replication.

We do not claim uniform gains across all metrics: the F2-2 vs. Cycle-local-4 success-rate difference is 1.53 percentage points (paired 95% CI [−1.153, 4.221] spans zero; Table A5), and 4-NFE recovery-rate intervals also span zero (Table A32). Two-NFE tests support lower response/recovery RMST, not recovery-rate gains (Tables A36, A38); runtimes are descriptive. Appendix A.1–A.3 and Table A1 specify route/seed matching, aggregating 220 routes per seed for inference across 8 paired seed-level outcomes (cycles are not independent replicates). Event inference preserves physical-event clusters; recovery uses paired cases stratified by seed and scenario.

6. Implementation and reporting.

Appendix P.1–P.2 gives the per-cycle procedure: each model generates its own history, episode boundaries reset the cache, and ground-truth targets never enter history. The selector picks the controller-facing trajectory; alignment uses realized ego state without assuming exact tracking. "Utility" is a geometric inheritance weight, not a value or safety probability. Equal geometric errors in Eq. 8 give label zero; invalid waypoints are masked from auxiliary losses, and unusable history triggers cycle-local fallback (Appendix B.3–B.4; Appendix P.1). The constructor adds 150,145 parameters (~0.096% of baseline); GPU memory is neither inferred from this ratio nor separately reported (Table A2; Appendix P.3). Table 1 separates published and controlled results.

Thank you again for your rigorous and constructive feedback, which has substantially strengthened our empirical analysis and helped clarify both our supported conclusions and remaining limitations.

AuthorsOpenReview ↗

Clarification of Novelty: A Scene-Adaptive Recurrent Generative Source

We would like to clarify the problem F2 addresses, its distinction from existing approaches, and how the experiments test its design.

The central problem is how to reuse a previous plan to reduce repeated generation without carrying forward stale intent. Only a short prefix of each planned trajectory is executed. The remaining future still contains useful decisions, but new observations and execution deviations can invalidate parts of it. At two NFEs, the refiner has limited opportunities to correct an unsuitable initialization. This motivates deciding, before refinement, which parts to retain and where to allow more exploration.

F2 advances and aligns the unexecuted future with actual execution, then predicts waypoint-wise inheritance weights using the current scene. These weights jointly adjust the source center and sampling spread: useful history is retained, while parts requiring revision move toward current candidates with broader sampling (§3.2, Eqs. 3–6).

Beyond previous-plan warm-starting, F2 learns how much history should contribute to the source center and scale at each waypoint. Relative to temporal denoising or anchor-to-trajectory generation, our focus is how execution and new observations update the recurrent source distribution for the next cycle. We claim this specific adaptive source construction as the methodological contribution; alignment and history reuse are not individually claimed as new.

The following controls test the practical role of this design:

  • Benefit beyond conventional reuse. Aligned warm-start receives the same execution-reconciled history and validity handling. F2 still improves DS by 0.42 at two NFEs, with a paired 95% CI of [0.130, 0.710].
  • Role of waypoint-wise center–spread adaptation. Fixed blending, center-only adaptation, fixed blending with adaptive spread, and scalar inheritance test the corresponding design choices. F2 outperforms these controls at both one and two NFEs, with the paired differences remaining significant after Holm correction (Table 2; Table A16).
  • Ability to revise incorrect history. In the two-NFE experiments, each method uses its own closed-loop history. F2 achieves lower all-case response or recovery RMST in both event suites and both fault suites, with all four paired confidence intervals below zero (Tables A35–A38). The additional Diffusion Planner/nuPlan evaluation tests the design in an independent system after adaptation and matched training. We hope this clarification makes both the technical contribution and its supporting evidence clearer. If it addresses your main concerns about novelty and applicability, we would be grateful if you would reconsider your current rating.

An aligned plan can still contain stale intent. The near-term path may remain useful even though a later maneuver needs to change. F2 uses the aligned plan, the current scene, and execution mismatch to estimate one inheritance weight at each waypoint. The same weight controls how far the source center shifts toward current candidates and how broadly the generator samples around that center. The contribution is this way of reconstructing the next cycle’s source distribution before refinement (§3.2, Eqs. 3–6).

The table reports mean Bench2Drive Driving Scores over eight matched training seeds. All variants use the same alignment and history-handling rules and two generator evaluations per cycle.

Configuration at two NFEsDriving Score ↑F2 − control (points)Paired 95% CIHolm-adjusted p
Aligned warm-start96.23+0.42[0.130, 0.710]0.02237
Fixed blend96.34+0.31[0.151, 0.469]0.02199
Adaptive center, fixed spread96.44+0.21[0.101, 0.319]0.02199
Fixed blend, adaptive spread96.40+0.25[0.108, 0.392]0.02199
Single inheritance weight for the full horizon96.50+0.15[0.058, 0.242]0.02237
F2 (joint waypoint-wise adaptation)96.65———

F2 improves on aligned warm-start and on each simpler source update in these tests (Table 2 and Table A16). The crossover helps explain when the source is useful. At two NFEs, substituting the F2 source for the ExecAlign source raises residual ADE from 0.142 to 0.177 m with the ExecAlign-trained refiner, while lowering it from 0.155 to 0.119 m with the F2-trained refiner (Tables A21–A22). On Diffusion Planner/nuPlan, we therefore evaluated F2 after adaptation and matched training. With the adaptation interface and two-NFE budget held fixed, F2 improves over aligned warm-start by 0.9–1.5 score points across all four settings, with each pointwise paired 95% interval above zero (Tables A39–A40).

Your concern about stale history also prompted the added two-NFE tests. F2 has lower response and recovery RMST than aligned warm-start in both event suites and both fault suites. Collision counts were 12/800 for F2 versus 19/800 for aligned warm-start in the abrupt-intent suite, and 6/160 versus 9/160 under sustained drift. These counts are descriptive. The revision also reports road and lane departures, TTC coverage, and cases of persistent inheritance or trajectory instability (Tables A35–A38 and A42–A49).

We hope these comparisons make it easier to assess what the design contributes and where the evidence applies. Thank you for the careful reading and for pointing us to the comparisons that needed to be clearer.

Reviewer_VUmPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes F2-Planner, which reduces redundant planning computation by aligning and adaptively reusing the unexecuted portion of the previous planning trajectory. On Bench2Drive, F2-Planner achieves a higher driving score than the four-call baseline using only two network calls, while reducing planning-head latency by 27.77%. The paper also evaluates its ability to revise plans through event-response and recovery-from-erroneous-history experiments under the four-call configuration.

Critical strengths

The paper addresses the practical issue of redundant computation in generative planning, presenting a clear approach to improving computational efficiency and providing reasonably comprehensive matched comparisons. Within the same planning system, F2-Planner achieves a DS of 96.65 with only two network calls, outperforming the cycle-local baseline that requires four calls, while reducing the average planning-head latency by 27.77%. These results demonstrate that effectively reusing the unexecuted portion of the previous planning trajectory can help reduce computational costs while maintaining strong planning performance.

Critical weaknesses

The main efficiency gains reported in the paper come from using only two network calls, whereas the event-response and recovery-from-erroneous-history experiments in §4.4 and Appendices H–I are all conducted with four calls. Since reducing the number of iterations also reduces the opportunities for the model to correct outdated plans, I am particularly interested in whether the two-call configuration can still respond promptly to changes. Providing corresponding results under this configuration would offer stronger evidence that the efficiency improvements do not come at the expense of adaptability to changing conditions.

The current results are primarily based on Bench2Drive v0.0.4. Although the additional experiments with a truncated diffusion generator are helpful, they still rely on the same perception, candidate generation, selection, and control system. Therefore, it remains difficult to determine how much the reported gains depend on this particular system. I would encourage the authors to further evaluate the method with other generative planners and provide results on additional datasets or closed-loop benchmarks to strengthen the evidence for its broader applicability.

Questions for authors

Are results on event response and recovery from erroneous history already available under the two-network-call configuration? If these experiments have not yet been conducted, could the authors use the existing event sets and failure settings to provide additional comparisons against the aligned warm-start baseline, reporting the restricted mean response time, recovery time, and the proportions of cases that fail to respond or recover? I would also appreciate it if each configuration could use the history generated by its own closed-loop rollouts, so that the evaluation captures potential error accumulation under different numbers of iterations. Beyond the two generative formulations currently evaluated, are there any validation results on other generators or datasets? If additional experiments are feasible, I would be particularly interested in whether the improvements over the aligned warm-start baseline remain consistent. Such results would help me better assess the generalizability and applicability of the proposed method.

3 comments
AuthorsOpenReview ↗

Response to Reviewer VUmP

Thank you for recognizing the practical motivation, matched experimental controls, and computational efficiency of our work. We greatly appreciate your constructive suggestions regarding low-budget responsiveness and cross-system generalizability. In the revised manuscript, we have incorporated new experiments that directly address both points.

1. Response and Recovery at 2 NFEs

We conducted additional event-response and corrupted-history recovery experiments at the 2-NFE operating point, retaining the original event definitions, response criteria, fault injections, and evaluation windows. Crucially, each configuration evolves using its own closed-loop history rather than borrowing cached history from the 4-NFE runs. Event identities and exogenous conditions are strictly matched across methods, while closed-loop ego states and recurrent histories evolve naturally per model. All evaluated methods use exactly two generator evaluations per cycle, with zero extra calls for F2 (Appendix K.1, p. 34).

The event-response results are reported in Tables A35 (p. 34) and A36 (p. 35). Compared against same-budget aligned warm-start (AWS), F2 reduces longitudinal and abrupt-intent response RMST by 39 ms and 25 ms, respectively, while corresponding non-response counts decrease from 21/800 to 6/800 and from 14/800 to 11/800. The corrupted-history recovery results are reported in Tables A37–A38 (p. 35). F2 reduces recovery RMST by 0.95 s under single-step faults and by 1.14 s under sustained drift, while unrecovered cases decrease from 7/160 to 4/160 and from 9/160 to 6/160, respectively.

All four RMST differences yield paired 95% confidence intervals strictly below zero. Our RMST computation accounts for all non-responding and unrecovered cases, utilizing paired cluster-bootstrap inference for event response and seed/scenario-stratified paired bootstrap for recovery (Appendix A.3, p. 15). While we do not claim statistical equivalence to 4 NFEs or statistically significant gains in overall response or recovery rates, these RMST reductions directly verify F2's responsiveness and recovery capability at the primary 2-NFE operating point.

2. Independent Planning System and Additional Closed-Loop Benchmark

We agree that substituting the generator within a single planning stack does not fully establish broader generalizability. We therefore integrated F2-Planner into Diffusion Planner on the nuPlan closed-loop benchmark, preserving its native scene encoder, joint prediction architecture, and control interface. Evaluations span Test14-random and Test14-hard under both non-reactive and reactive traffic. All controlled baselines operate at 2 NFEs with matched training seeds and evaluation protocols, sharing identical training data, training budgets, post-processing, and controllers (Appendix L.1, pp. 35–36).

Table A39 (p. 36) presents the complete comparisons across the native Diffusion Planner architecture and controls sharing a common adaptation interface. As shown in Table A40 (p. 36), F2 improves the nuPlan closed-loop score over AWS by 0.9–1.5 points on the 100-point scale across all four settings, with every pointwise paired 95% confidence interval strictly above zero.

These experiments extend our empirical validation to an independent planning architecture and closed-loop benchmark. Taken together with our source–refiner crossover analysis (Appendix G, Tables A21–A22, pp. 25–26), these results support the applicability of the proposed source update when paired with interface adaptation and matched source–refiner training, rather than claiming training-free transfer to arbitrary planners.

We have also summarized these findings in Section 4.4 of the revised manuscript. Thank you again for your valuable feedback, which has helped strengthen the empirical evidence for F2-Planner. Should any points warrant further clarification, we remain readily available and would greatly appreciate your timely response while the discussion period remains open.

Thank you for your response and the additional experiments. I have noted the updates and appreciate your efforts in addressing my comments.

Thank you for your time and thoughtful feedback. We appreciate your consideration of our response and additional experiments, and your suggestions have helped us strengthen the paper.

Comments on the paper

1 comment
Area_Chair_U2fZOpenReview ↗

Discussion

Dear Reviewers, the authors have posted their rebuttal. Please review their responses and engage in the discussion.

Best, AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.