acceptodds
Under review as a conference paper at ICLR 2027

Learning When to Search: Adaptive Test-Time Computation for Generative Motion Planning

Abstract

In autonomous driving, generative motion planners often use candidate generation and scoring to improve the final trajectory, but running this search process at every planning step introduces additional computation. In this study, we investigate when search is actually needed. Our analysis shows that the benefit of search varies greatly across planning steps and is related to both the current planning information and recent history. We therefore develop a lightweight history-aware model to decide whether search should be performed after the base trajectory is generated. If search is needed, the original search module generates and evaluates additional candidates; otherwise, the base trajectory is used directly. Closed-loop experiments on nuPlan show that our method achieves an average score of 86.07 with search on 56.81% of planning steps, comparable to Full Search at 85.64 with 100% calls. It also performs better than fixed search schedules at similar search rates and reduces mean planning inference time by 32.0%.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Motion Planning, Generative Models, Adaptive Test-Time Computation, trajectory planning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TNLfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a motion planner that decides each step if trajectory search among candidates is necessary or not. In closed-loop experiments it shows that the approach can save search in almost half of the steps. For some planners, such as the Diffusion Planner, it can even improve performance compared to doing a search at every step. The paper contributes also investigations into benefits of search and search frequency scaling on the Val14 Development Set of nuPlan. Decisions are learned from a created dataset of paired trajectory alternatives for each step, one from the base planner and one which is the result of a search in that step. With this, a small score prediction model is trained to decide if a search should be done based on the most recent planning steps and a custom set score threshold.

Critical strengths

Given the idea of re-planning is investigated in the robotic domain for decades, this application in autonomous driving is surprisingly simple and efficient. Predicting the need for trajectory search based on PDM component scores rather than inside of a full RL pipeline should allow for fast training and produces surprisingly good results.

It is positive that the method achieves on-par performance with the diffusion planner with full PDM search but at almost half the amount of searches. In general, the reduction of search is significant even if the performance of other planners sometimes suffers. The trade-off seems minor.

Independent of the method and results, the investigation uncovered that, given the nuPlan dataset and Diffusion planner, the top 20% of the planning steps account for 78.3% of the total positive improvement. This is an interesting insight for the domain.

The approach works as added module for any planner with a compatible search step.

Critical weaknesses

The related work is a bit light on re-planning approaches and the novelty of treating it as a "when to search" task is smaller than it may appear from the phrasing. Apart from that, relevant close work such as [1] which investigated deciding on when to replan with reinforcement learning should be cited.

While the investigation into search benefits seems solid, a 0,164 and 0.277 Spearman correlation of route heading changes and heading differences to the base trajectory is rather low and concluding that these relations provide useful clues seems optimistic.

The runtime analysis does not look into peak planning conditions. The analysis shows a 32% reduction in planning time but this is averaged over the whole dataset. In practice, this may serve to safe searches overall but in situations where every step does need a re-planning step, the search will only add overhead. This is especially important when the expectation of some challenges is to provide trajectories with 10 fps. There may be a conceptual issue in this: Why use the method if it saves planning steps overall but adds to peak computation and inference time? This is not a major issue since only the fast predictor runtime is added but this may affect the overall impact.

Simple baselines to compare against would be useful, e.g. one that decides randomly to search or not and another history based baseline that searches if the recent history required searches and vice versa. Such a baseline can be expected to perform well if there is such a strong temporal correlation of the need for search.

While it is clever to use PDM scores for the search decision this is different from comparing future closed-loop rollouts as the paper admits in A.2.4. PDM scoring and the true objective may not be aligned in a systematic way and the approach will inherit PDM score issues.

The evaluated planners and improved planners in a closed loop setting are necessarily tested on separate state distributions because their trajectories should drift apart. It would be useful to evaluate for the visited states of the base planner if search would have been executed. That way the same states would have been visited, even though in an offline way, but which would allow for a comparison on the same state distribution. This is similar to 5.3 but not the same.

The repeated call control seems to have an outsized effect on the final result and should be mentioned in the final paper. A lot of similar details in the appendix make the approach appear a bit engineered.

[1] Honda, K., Yonetani, R., Nishimura, M., & Kozuno, T. (2024, May). When to replan? an adaptive replanning strategy for autonomous navigation using deep reinforcement learning. In 2024 IEEE International Conference on Robotics and Automation (ICRA) (pp. 6650-6656). IEEE.

Questions for authors

Is there any insight why the score increase vs. search rate is not strictly monotonic?

Could you give any indication on the variance of the closed-loop score over different runs with different random seeds?

Could you describe the 4 of 5 rule and other details in repeated call control more closely? Was this estimated? How robust are these results to using more or less than 8 positive requests?

Without clarification of the issues described in critical weaknesses and answers to these questions I see the paper as borderline, slightly below the acceptance threshold but given a solid rebuttal I see potential to give a significantly more positive rating.

7 comments
AuthorsOpenReview ↗

Response to TNLf (1/5): Related work and feature choice

Thank you for carefully reviewing our work and for your specific and insightful suggestions. Following your comments, we have added methodological explanations and experimental comparisons. In particular, we further validated direct threshold-based decisions after removing the hand-crafted repeated-call rules and adopted this as our revised main method.

We refer to the method in the original submission as Submitted Ours. In addition to learning search value, it uses hand-crafted rules to control repeated calls. At each frame, the base planner first generates a new trajectory. The method then uses current and historical information to predict PDM component scores, component changes due to search, and degradation probabilities, combining them into a search value vv. Search is executed only when vv exceeds the threshold δ\delta and passes the repeated-call control. Inspired by your suggestion, we removed the hand-crafted repeated-call control rules and refer to the simplified method as Revised Ours. During its mature phase, it decides whether to search directly using v>δv > \delta: CALL executes PDM candidate search, whereas SKIP executes the current frame's base trajectory.

Critical Weakness 1

The related work is a bit light on re-planning approaches and the novelty of treating it as a "when to search" task is smaller than it may appear from the phrasing. Apart from that, relevant close work such as [1] which investigated deciding on when to replan with reinforcement learning should be cited.

Thank you for recommending Honda et al.'s When to Replan? An Adaptive Replanning Strategy for Autonomous Navigation Using Deep Reinforcement Learning (ICRA 2024) [1]. We will add this reference and precisely define our contribution. That work learns when to regenerate a global reference path in hierarchical navigation; in our method, a new base trajectory is generated at every step, and the learned module decides whether to add candidate search. Our contribution lies in analyzing the distribution of benefits from this additional search and allocating its computation using a lightweight supervised model.

[1] Honda, K., Yonetani, R., Nishimura, M., & Kozuno, T. (2024, May). When to replan? an adaptive replanning strategy for autonomous navigation using deep reinforcement learning. In 2024 IEEE International Conference on Robotics and Automation (ICRA) (pp. 6650-6656). IEEE. Full text.

Critical Weakness 2

While the investigation into search benefits seems solid, a 0,164 and 0.277 Spearman correlation of route heading changes and heading differences to the base trajectory is rather low and concluding that these relations provide useful clues seems optimistic.

Thank you for your suggestion. We added a direct closed-loop comparison: the heading-deviation threshold policy scores 0.672 points higher than Random while using fewer search calls, showing that this feature contains information useful for search decisions. Below are the mean results on the Random-NR, Random-R, Hard-NR, and Hard-R test sets with test-run seed 0.

MethodDriving score / 100Mature-frame call rate
Revised Ours86.59155.44%
Random84.88256.81%
Heading-deviation threshold85.55453.76%

Route heading change describes the need to turn ahead, while the heading deviation between the base trajectory and the route describes their directional consistency. Our search predictor jointly models these quantities with the traffic environment, route geometry, and base-trajectory dynamics, using 25 factors in total and recent history. Revised Ours achieves 86.591 points, a further 1.037-point improvement over the heading-deviation threshold policy alone.

AuthorsOpenReview ↗

Response to TNLf (2/5): Search runtime and peak computation

Critical Weakness 3

The runtime analysis does not look into peak planning conditions. The analysis shows a 32% reduction in planning time but this is averaged over the whole dataset. In practice, this may serve to safe searches overall but in situations where every step does need a re-planning step, the search will only add overhead. This is especially important when the expectation of some challenges is to provide trajectories with 10 fps. There may be a conceptual issue in this: Why use the method if it saves planning steps overall but adds to peak computation and inference time? This is not a major issue since only the fast predictor runtime is added but this may affect the overall impact.

Thank you for highlighting the importance of peak computation under 10 Hz planning. We substantially reduced the computation time of each search through implementation optimizations that reuse repeated route encodings, reduce GPU kernel dispatch, cache PDM route geometry, and batch projection calculations. After optimization, CALL mean runtime is 57.56 ms and P95 is 71.53 ms, both below the 100 ms planning interval, easing latency pressure on search frames.

Revised Ours completed 1,066 tests across the four test sets on RTX 4090 with test-run seed 0. These engineering optimizations preserve the original model weights, FP32 precision, and 21 candidates. The mean closed-loop score remains nearly unchanged after optimization: 86.521 before versus 86.553 after (+0.033 points). The search predictor takes only 4.14 ms on average; on-demand search reduces total computation, while implementation optimization further reduces runtime when search is needed.

Model (ms)Search predictor (ms)PDM (ms)Other overhead (ms)Core total (ms)
13.964.1433.875.5957.56

These are mean runtimes over all CALL frames. The search predictor includes feature extraction, processing, and prediction; other overhead includes trajectory conversion, search wrapping, and the remaining call overhead.

Test setCALL core mean (ms)CALL core P95 (ms)Over 100 ms
Hard-NR58.7972.651.00%
Hard-R57.5071.290.95%
Random-NR57.1571.340.95%
Random-R56.5270.200.91%

The mean core runtime over all decision frames is 42.37 ms. Of these, 926 / 149,400 frames exceed 100 ms (0.6198%); among CALL frames, the corresponding count is 794 / 82,857 (0.9583%).

StatisticDriving score / 100Mature-frame call rate
Mean after optimization86.55355.37%
AuthorsOpenReview ↗

Response to TNLf (3/5): Simple baselines and PDM supervision

Critical Weakness 4

Simple baselines to compare against would be useful, e.g. one that decides randomly to search or not and another history based baseline that searches if the recent history required searches and vice versa. Such a baseline can be expected to perform well if there is such a strong temporal correlation of the need for search.

Thank you for suggesting these two direct comparisons. Additional experiments on the Random-NR, Random-R, Hard-NR, and Hard-R test sets show that Revised Ours has a clearly higher macro-average driving score than Random and History: it achieves 86.591 points with a 55.44% call rate, exceeding History by 1.486 points and Random by 1.710 points. Submitted Ours also scores higher than these simple rules. The table reports the equally weighted macro-average over Random-NR, Random-R, Hard-NR, and Hard-R (261/272 scenarios in Random/Hard, respectively) with test-run seed 0, together with each method's actual call rate.

MethodDriving score / 100Mature-frame call rate
Revised Ours86.59155.44%
Submitted Ours86.07456.81%
Full Search85.685100.00%
History85.10551.79%
Random84.88256.81%

History continues searching when the most recently observed search gain exceeds 0.05; otherwise, it probes after six planner calls have elapsed since the last search. The probing interval is fixed on Val14. Compared with Submitted Ours, the direct threshold-based decisions of Revised Ours further improve the driving score and reduce the call rate.

Critical Weakness 5

While it is clever to use PDM scores for the search decision this is different from comparing future closed-loop rollouts as the paper admits in A.2.4. PDM scoring and the true objective may not be aligned in a systematic way and the approach will inherit PDM score issues.

Thank you for your comment. The scenario-average post-search PDM score over Revised Ours' actual CALL steps has a Spearman rank correlation of 0.456 with its final closed-loop score. This shows a positive association between local PDM scores and final closed-loop driving performance.

PDM's sub-objectives are themselves constructed around closed-loop driving evaluation. In Section 3.3 of Parting with Misconceptions about Learning-based Vehicle Motion Planning (CoRL 2023) [1], Dauner et al. explicitly state that the scoring function "closely resembles the nuPlan evaluation metrics". Candidate execution is first simulated using an LQR controller and a bicycle model with the same parameters as nuPlan, and candidates are then scored for traffic compliance, progress, and comfort.

We use the StateTransformer/PDM refinement branch integrated into Diffusion Planner [3]. The supervised components include no at-fault collision (NC), time to collision (TTC), drivable area compliance (DAC), driving direction compliance (DDC), progress (P), comfort (C), and speed limit compliance (SL). These components incorporate the safety, rule compliance, efficiency, and comfort objectives of final closed-loop evaluation into short-horizon candidate evaluation. Diffusion Planner [3] and NAVSIM [2] also provide evidence of closed-loop improvements from refinement and an association between PDM scores and closed-loop performance, respectively.

With this supervision, Revised Ours improves the average score over Base by 7.555 points and over Full Search by 0.907 points under test-run seed 0. Across three test-run seeds, its mean complete closed-loop score is 86.429, compared with 85.469 for Full; their search call rates are 55.59% and 100%, respectively.

[1] Dauner et al. Parting with Misconceptions about Learning-based Vehicle Motion Planning. CoRL 2023. Official paper.

[2] Dauner et al. NAVSIM: Data-Driven Non-Reactive Autonomous Vehicle Simulation and Benchmarking. NeurIPS 2024. Full text.

[3] Zheng et al. Diffusion-Based Planning for Autonomous Driving with Flexible Guidance. ICLR 2025. Full text.

Reviewer_DyhgOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a history-aware predictor that decides when to invoke candidate search after base trajectory generation. On nuPlan, it achieves comparable average performance to Full Search with 56.81% search calls and 32% lower mean planning time. Experiments cover four planners and ablate the main components.

Critical strengths
  • The method offers a practical computation–performance trade-off, supported by controlled comparisons and runtime measurements.

  • Targeted ablations and experiments across multiple planners support the design and its applicability.

Critical weaknesses
  • The evaluation emphasizes mean runtime savings rather than reliable operation within a planning deadline. Skipping search in easy scenes does not establish that search remains timely in complex scenes.

  • Synchronous simulation does not capture the effect of planning latency on driving. The runtime savings are demonstrated, but their benefit under real-time execution remains unvalidated.

  • The predictor controls whether search is skipped, but its reliability in retaining safety-critical search interventions is not established. Average scores and local-gain coverage do not reveal consequential missed interventions.

Questions for authors
  • Can you report collision, TTC, progress, and comfort metrics to explain the efficiency-safety trade-off in reactive setting?

  • Are the predictor’s training pairs collected separately for each base planner, and is the predictor retrained for each? Please also clarify how the 606 training scenarios and the calibration set are selected and separated.

  • Could you report the P95 planning latency for steps that invoke search, alongside the mean latency and the fraction exceeding the 100 ms planning interval?

4 comments
AuthorsOpenReview ↗

Response to Dyhg (1/4): Search implementation and planning deadlines

Thank you for carefully reviewing our work and for recognizing its trade-off between computation and driving performance. Following your suggestions on search runtime, driving under latency, and safety interventions, we have added the corresponding experiments, which we address individually below.

We refer to the method in the original submission as Submitted Ours. In addition to learning search value, it uses hand-crafted rules to control repeated calls. At each frame, the base planner first generates a new trajectory. The method then uses current and historical information to predict PDM component scores, component changes due to search, and degradation probabilities, combining them into a search value vv. Search is executed only when vv exceeds the threshold δ\delta and passes the repeated-call control. Inspired by Reviewer TNLf's suggestion, we removed the hand-crafted repeated-call control rules and refer to the simplified method as Revised Ours. During its mature phase, it decides whether to search directly using v>δv > \delta: CALL executes PDM candidate search, whereas SKIP executes the current frame's base trajectory.

Critical Weakness 1

The evaluation emphasizes mean runtime savings rather than reliable operation within a planning deadline. Skipping search in easy scenes does not establish that search remains timely in complex scenes.

Thank you for your suggestion. Inspired by your concerns about planning deadlines and peak computation in complex scenarios, we further optimized the search implementation. We use a CUDA Graph for the entire solver, reuse repeated route encodings, and reduce GPU kernel dispatch. We also use array-based geometry optimization to cache PDM route geometry and batch projection calculations. These changes substantially reduce the computation time of each search, helping search finish on time in complex scenarios.

Revised Ours completed 1,066 tests across the four test sets on RTX 4090 with test-run seed 0. These engineering optimizations preserve the original model weights, FP32 precision, and 21 candidates. The mean closed-loop score remains nearly unchanged after optimization: 86.521 before versus 86.553 after (+0.033 points). The table below breaks down mean runtime on CALL frames.

Model (ms)Search predictor (ms)PDM (ms)Other overhead (ms)Core total (ms)
13.964.1433.875.5957.56

These are mean runtimes over all CALL frames. The search predictor includes feature extraction, processing, and prediction; other overhead includes trajectory conversion, search wrapping, and the remaining call overhead.

StatisticDriving score / 100Mature-frame call rate
Mean after optimization86.55355.37%
AuthorsOpenReview ↗

Response to Dyhg (2/4): Closed-loop results with 100 ms timeout fallback

Critical Weakness 2

Synchronous simulation does not capture the effect of planning latency on driving. The runtime savings are demonstrated, but their benefit under real-time execution remains unvalidated.

Thank you for your suggestion. The official nuPlan challenge uses a simulation frequency of 10 Hz, corresponding to 100 ms per step. Accordingly, we completed a 100 ms planning-deadline test of Revised Ours on RTX 4090. Specifically, at each frame, we check whether the current planning output is ready within 100 ms: if it is on time, we use the original synchronous output; if it exceeds 100 ms, we do not use that frame's late planning output and instead prioritize the current frame's Base trajectory if it is ready by the deadline. If Base is also not ready, we retain the previous valid trajectory. Thus, fallback replaces the executed trajectory only on late frames, while on-time frames retain their original synchronous behavior.

After further implementation optimization, we used the full timing measurements to complete fallback tests for the 559 scenario runs that experienced timeouts, retaining the original results for the other 507 runs. The table below reports the mean driving score (out of 100) across Hard-NR, Hard-R, Random-NR, and Random-R after combining all 1,066 runs, with test-run seed 0.

ScopeOriginal synchronousTimeout fallbackDifference
Mean over 4 test sets (1,066 runs combined)86.55386.155-0.398

Under the 100 ms deadline rule, Revised Ours has an average score change of -0.398 points relative to the original synchronous run. After fallback, it still achieves 86.155 points, which is 0.471 points higher than the synchronous Full Search score of 85.685.

Meanwhile, Revised Ours has a clearly lower search call rate: this configuration uses 55.37% in the complete synchronous tests, 44.63 percentage points below Full Search's 100%. PDM search averages 33.87 ms and is the main computational overhead on CALL frames. Full executes search at every frame and therefore more frequently incurs the pressure on the 100 ms deadline from this additional computation; Revised Ours reduces these expensive steps through on-demand calls.

AuthorsOpenReview ↗

Response to Dyhg (3/4): Safety-critical interventions and reactive driving metrics

Critical Weakness 3

The predictor controls whether search is skipped, but its reliability in retaining safety-critical search interventions is not established. Average scores and local-gain coverage do not reveal consequential missed interventions.

Thank you for suggesting that we examine search interventions with actual safety consequences. We added an analysis of officially evaluated risk periods and a closed-loop safety comparison between Base and Revised Ours for all 1,066 runs on the Hard-NR, Hard-R, Random-NR, and Random-R test sets, with test-run seed 0.

On fixed Base driving states, the predictor's search request rate on frames identified as TTC risks by the official evaluation is 83.10%, higher than the 55.20% rate over all available decision frames. Among 83 runs with complete records for the 2 seconds preceding the first at-fault collision, 82 request search in advance, a coverage rate of 98.80%.

Of the 101 runs in which Base has at-fault collisions, Revised Ours eliminates at-fault collisions in 77 runs and reduces the number of collision events in another 3. Thus, at-fault collisions are reduced in 80/101 runs (79.21%). Across the full test set, at-fault collision events decrease from 158 to 34, a reduction of 78.48%.

Among these 80 runs with fewer collisions, using Base's first collision time as the reference, 73 have complete records for the preceding 2 seconds. Revised Ours actually invokes search in its own closed-loop run in 71 of these cases (97.26%), and the fraction of search frames within these windows is 82.89%.

Question 1

Can you report collision, TTC, progress, and comfort metrics to explain the efficiency-safety trade-off in reactive setting?

Thank you for suggesting that we report individual driving components. Under reactive interactions, maintaining safety with fewer search calls is more challenging; Revised Ours achieves driving scores close to Full with clearly fewer calls. Other traffic participants respond to the ego vehicle's actions, making it necessary to examine call rates, total scores, and individual driving components together.

Averaged over three test-run seeds, the call rates on Hard-R and Random-R are 59.07% and 52.34%, respectively, reductions of 40.93 and 47.66 percentage points from Full's 100%. The corresponding driving scores are 79.697 and 90.448, compared with 80.031 and 91.892 for Full, with gaps of 0.334 and 1.444 points, respectively.

Test setMethodNC ↑TTC ↑Progress ↑Comfort ↑Events ↓ s0/s1/s2
Hard-RRevised Ours95.3487.3878.4086.0315/17/19
Hard-RFull97.3087.6278.7784.688/10/10
Random-RRevised Ours97.7094.0088.2589.146/6/6
Random-RFull98.3494.2590.2389.404/5/4

Revised Ours has a score advantage on the other test sets and in the overall mean. With the same three seeds, its scores on Hard-NR and Random-NR are 80.491 and 95.080, exceeding Full by 3.240 and 2.379 points, respectively. The equally weighted mean over the four test sets is 86.429, compared with 85.469 for Full, an improvement of 0.960 points, with a mean call rate of 55.59%.

AuthorsOpenReview ↗

Response to Dyhg (4/4): Training-data separation and CALL latency

Question 2

Are the predictor’s training pairs collected separately for each base planner, and is the predictor retrained for each? Please also clarify how the 606 training scenarios and the calibration set are selected and separated.

Thank you for your question.

All base planners share the same 25-dimensional input feature definitions, search predictor architecture, and learning procedure; the training data, predictor parameters, and threshold are adapted separately to each base planner. Different base planners generate different trajectories, visit different states, and obtain different search benefits in the same Val14 scenarios. We therefore collect Base/Search pairs separately for each base planner, train a predictor with the same architecture, and calibrate its threshold independently. The predictor thus learns the search value for that base planner. We will clarify this procedure in the revision.

We partition the 1,118 Val14 scenarios using a fixed hash ordering of their unique scenario identifiers: the first 256 form the internal diagnostic set, the next 256 form the calibration set, and the remaining 606 are used for supervised training of the search predictor. This partition is determined before training and calibration, and all frames from a scenario always belong to the same group.

The three groups have disjoint scenario identifiers; external Test14 and Val14 have no overlapping scenario identifiers or logs. The MiniRocket representation is fitted without supervision on all of Val14 and then frozen. Supervised training of the search predictor uses only the 606 training scenarios; calibration scenarios do not participate in supervised parameter updates.

Once selected, the threshold remains fixed across all four test sets and all test-run seeds. When recollecting training data for different base planners, the scenario partitions for training, calibration, and internal diagnostics remain unchanged.

Revised Ours retains the original search predictor and component-level supervision, changing only execution control in the mature phase and the threshold.

Question 3

Could you report the P95 planning latency for steps that invoke search, alongside the mean latency and the fraction exceeding the 100 ms planning interval?

Thank you for suggesting that we report the mean runtime, P95, and fraction exceeding 100 ms for search frames. Below are the results from 1,066 tests of Revised Ours across the four test sets on RTX 4090, with test-run seed 0.

Test setCALL core mean (ms)CALL core P95 (ms)Over 100 ms
Hard-NR58.7972.651.00%
Hard-R57.5071.290.95%
Random-NR57.1571.340.95%
Random-R56.5270.200.91%

The mean core runtime over all decision frames is 42.37 ms. Of these, 926 / 149,400 frames exceed 100 ms (0.6198%); among CALL frames, the corresponding count is 794 / 82,857 (0.9583%).

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

The paper investigates when the search process is actually needed in autonomous planning. The test-time analysis proves the searching effecitveness varies in different planning steps and also in current/history ones. A history-induced module is designed to decide whether to search after base planning. They prove the effectiveness on nuPlan closed-loop experiments.

Critical strengths
  1. The overall analysis seems to be reasonable. The paper first analyzes the benefits of search, search frequency, and influence on current/history context.

  2. The paper's writing is easy to follow. Although I have little priors and expertise, I can still get the main meanings from the paper.

  3. The paper not only proves the effectiveness on one planner, but also generalizes to different baselines.

Critical weaknesses
  1. It is still unclear about the relationship between local search benefit and long-horizon closed-loop benefits. The paper curates the supervision signals by score changes of PDM. However, as proved in Appendix A2.4, the difference cannot reflect the real closed loop benefit differences.

  2. In Table 1, full search achieves 85.64, with only 0.43 lower than the proposed method. The authors fail to provide the multi-run sensitiveness and flunctuation range across runs. This limited improvement cannot be directly attributed to the designed search.

  3. The comparison baselines are somewhat simplified. The method is mainly compared with periodic search, but the paper itself identifies route heading changes and route heading deviations as useful predictors. More comparisons should be added, like simple threshold-based policies using these features, random search at a similar budget, and lightweight classifiers.

Questions for authors

In the generalization experiment, for different baselines, do they use the same predictor and threshold, or retrained? This can reveal whether the strategy can be directly transferred.

5 comments
AuthorsOpenReview ↗

Response to QVLn (1/3): Local PDM supervision and closed-loop performance

Thank you for carefully reviewing our work and for recognizing our analysis and cross-planner experiments. Following your suggestions on local supervision, repeated runs, and simple baselines, we have added the corresponding experiments. Inspired by Reviewer TNLf's suggestion, we have also removed the hand-crafted repeated-call rules.

We refer to the method in the original submission as Submitted Ours. In addition to learning search value, it uses hand-crafted rules to control repeated calls. At each frame, the base planner first generates a new trajectory. The method then uses current and historical information to predict PDM component scores, component changes due to search, and degradation probabilities, combining them into a search value vv. Search is executed only when vv exceeds the threshold δ\delta and passes the repeated-call control. Inspired by Reviewer TNLf's suggestion, we removed the hand-crafted repeated-call control rules and refer to the simplified method as Revised Ours. During its mature phase, it decides whether to search directly using v>δv > \delta: CALL executes PDM candidate search, whereas SKIP executes the current frame's base trajectory.

Critical Weakness 1

It is still unclear about the relationship between local search benefit and long-horizon closed-loop benefits. The paper curates the supervision signals by score changes of PDM. However, as proved in Appendix A2.4, the difference cannot reflect the real closed loop benefit differences.

Thank you for your comment. The scenario-average post-search PDM score over Revised Ours' actual CALL steps has a Spearman rank correlation of 0.456 with its final closed-loop score. This shows a positive association between local PDM scores and final closed-loop driving performance.

PDM's sub-objectives are themselves constructed around closed-loop driving evaluation. In Section 3.3 of Parting with Misconceptions about Learning-based Vehicle Motion Planning (CoRL 2023) [1], Dauner et al. explicitly state that the scoring function "closely resembles the nuPlan evaluation metrics". Candidate execution is first simulated using an LQR controller and a bicycle model with the same parameters as nuPlan, and candidates are then scored for traffic compliance, progress, and comfort.

We use the StateTransformer/PDM refinement branch integrated into Diffusion Planner [3]. The supervised components include no at-fault collision (NC), time to collision (TTC), drivable area compliance (DAC), driving direction compliance (DDC), progress (P), comfort (C), and speed limit compliance (SL). These components incorporate the safety, rule compliance, efficiency, and comfort objectives of final closed-loop evaluation into short-horizon candidate evaluation. Diffusion Planner [3] and NAVSIM [2] also provide evidence of closed-loop improvements from refinement and an association between PDM scores and closed-loop performance, respectively.

With this supervision, Revised Ours improves the average score over Base by 7.555 points and over Full Search by 0.907 points under test-run seed 0. Across three test-run seeds, its mean complete closed-loop score is 86.429, compared with 85.469 for Full; their search call rates are 55.59% and 100%, respectively.

[1] Dauner et al. Parting with Misconceptions about Learning-based Vehicle Motion Planning. CoRL 2023. Official paper.

[2] Dauner et al. NAVSIM: Data-Driven Non-Reactive Autonomous Vehicle Simulation and Benchmarking. NeurIPS 2024. Full text.

[3] Zheng et al. Diffusion-Based Planning for Autonomous Driving with Flexible Guidance. ICLR 2025. Full text.

AuthorsOpenReview ↗

Response to QVLn (2/3): Stability across test-run seeds

Critical Weakness 2

In Table 1, full search achieves 85.64, with only 0.43 lower than the proposed method. The authors fail to provide the multi-run sensitiveness and flunctuation range across runs. This limited improvement cannot be directly attributed to the designed search.

Thank you for your suggestion. We fixed the search predictor and threshold and added complete closed-loop experiments with three test-run seeds.

Test-run seed: with the model fixed, assess whether driving performance is stable under different trajectory samples.

Submitted Ours still has a higher three-seed mean than Full, while executing search on only 56.86% of frames. With Revised Ours, inspired by TNLf's suggestion, the call rate is 55.59%, and the mean score exceeds Full by 0.960 points, approximately 1 point. The following results cover the Random-NR, Random-R, Hard-NR, and Hard-R test sets, totaling 3,198 runs per method.

MethodMean ± SD across three test-run seedsMature-frame call rate
Revised Ours86.429 ± 0.26355.59%
Submitted Ours85.743 ± 0.33156.86%
Full Search85.469 ± 0.215100.00%
Test-run seedRevised OursSubmitted OursFull
086.59186.07485.685
186.57085.74185.254
286.12685.41385.468

The test-run seeds are 0/1/2. The Full column reports the measured results for each test-run seed; Revised Ours scores +0.907/+1.316/+0.658 points higher, respectively.

AuthorsOpenReview ↗

Response to QVLn (3/3): Additional baselines and adaptation to base planners

Critical Weakness 3

The comparison baselines are somewhat simplified. The method is mainly compared with periodic search, but the paper itself identifies route heading changes and route heading deviations as useful predictors. More comparisons should be added, like simple threshold-based policies using these features, random search at a similar budget, and lightweight classifiers.

Thank you for suggesting these direct comparisons. The additional experiments show that Revised Ours achieves a clearly higher driving score than these simple policies: it exceeds the strongest of them, heading deviation, by 1.037 points, logistic by 1.733 points while using fewer calls, and Random by 1.710 points. Below are equally weighted macro-averages over the Random-NR, Random-R, Hard-NR, and Hard-R test sets (261/272 scenarios in Random/Hard, respectively) with test-run seed 0, together with the actual call rates.

MethodDriving score / 100Mature-frame call rate
Revised Ours86.59155.44%
Submitted Ours86.07456.81%
Full Search85.685100.00%
Random84.88256.81%
Heading-change threshold84.81755.72%
Heading-deviation threshold85.55453.76%
Current-feature logistic84.85861.21%

Revised Ours achieves 86.591 points with a 55.44% call rate; heading deviation achieves 85.554 points with 53.76%, and logistic achieves 84.858 points with 61.21%. Around these actual call rates, history-conditioned component prediction provides a better combination of driving score and computation.

Question 1

In the generalization experiment, for different baselines, do they use the same predictor and threshold, or retrained? This can reveal whether the strategy can be directly transferred.

Thank you for your question. The search predictor is trained separately for each base planner, and its threshold is calibrated independently.

All base planners share the same 25-dimensional input feature definitions, search predictor architecture, and learning procedure; the training data, predictor parameters, and threshold are adapted separately to each base planner. Different base planners generate different trajectories, visit different states, and obtain different search benefits in the same Val14 scenarios. We therefore collect Base/Search pairs separately for each base planner, train a predictor with the same architecture, and calibrate its threshold independently. The predictor thus learns the search value for that base planner. We will clarify this procedure in the revision.

Training data are recollected through simulation for each base planner, but the scenario partitions for training, calibration, and internal diagnostics remain identical across base planners.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.