- Rating
- 3(+1)
- Confidence
- 2(−1)
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
CLOVER addresses a critical issue: simply imitating a single human trajectory during training does not equate to learning to generate a variety of safe, feasible behaviors with high planning scores. The method first expands the candidate coverage using pseudo-expert trajectories selected by an evaluator; it then alternately trains the scorer and uses the top-k or Pareto trajectories selected by the scorer to supervise the generator. The contribution lies in first providing the planner with more valid options, and then conservatively utilizing an imperfect scorer to improve the candidate set and the final selection.
Critical strengths
The paper keeps the original backbone and improves DrivoR's score on NAVSIM v1 by adjusting the training method without introducing any other information. The analysis of candidate trajectories is also quite thorough. The results show that Stage 1 does indeed improve the generated candidates, and it's not just about the scorer being more selective. Table 8 also illustrates the necessity of alternating updates: synchronous updates of the generator and scorer will crash, while alternating updates maintain stability.
Critical weaknesses
The result on NAVSIM v2 (EPDMS 90.4) requires further clarification. The core model achieves only 89.3, the 90.4 score relies on an additional, monomodal anchor trained to suppress trajectories that deviate too far from the 64 candidates. However, the paper does not clearly explain how the anchor’s weights were determined. Appendix D states that the search was conducted on a “validation set,” but NAVSIM does not have a commonly used validation set independent of navtest; therefore, it must be clearly specified which partition was used here. The role of the anchor also highlights a more direct issue: without the anchor, the EC is only 35.0, while other methods all exceed 75. The explanation provided in the paper is that the model generates a set of candidates with similar scores but different behaviors; however, since the highest-scoring candidate is selected independently for each frame, the model might move left in one frame and right in the next. This explanation does not resolve the issue; rather, it highlights the lack of continuity in frame-by-frame selection. A method named “closed-loop” yet clearly lagging behind existing methods on the very metric directly related to continuous decision-making is, in my view, a shortcoming that needs to be addressed head-on.
The differences between the two stages and existing methods have not yet been clearly explained. Stage 1 enumerates lateral offsets, speed, and acceleration/deceleration in the lane coordinate system, filters them using the PDM evaluator, and then uses a min-of-K loss to have the generator cover these “pseudo-expert” trajectories. Scoring by a rule-based teacher and then distilling the results to the model is already a practice adopted by the Hydra-MDP series; the change here is replacing a fixed vocabulary with network-generated candidates. The paper lacks a comparison using “the same backbone, the same scorer, and vocabulary distillation,” so it is impossible to determine whether the improvement stems from the learned candidates or from the supervision provided by the PDM. As I understand it, Stage 2 involves having the learned scorer select trajectories, with the generator fitting these trajectories while being constrained not to deviate too far from the previous version, and then refitting the scorer. This process is essentially consistent with methods such as expert iteration, ReST, and rejection-sampling fine-tuning, except that it is applied to trajectory generation. The paper devotes an entire section and a figure to emphasize that “this is not RL,” but what requires further explanation is: how exactly does it differ from “one-step offline RL combined with a learned reward” in terms of specific training objectives and update mechanisms?
Questions for authors
-
On which dataset was the parameter search in Table 13 performed?
-
Were anchors used in the NavHard results in Table 4? Please list both versions of EPDMS and EC with and without anchors.
-
What is the inference latency with anchors? The 110 ms mentioned in the text refers to the case without anchors, correct?
-
In most NAVSIM scenarios, the difference is observed only in EP. Does this mean the Pareto set degenerates into “the top EP”? Could this explain why EP outperforms humans and why EC collapses when anchors are not used?
-
Eq. 97 breaks down the top-1 improvement into two components: an increase in the candidate pool upper bound and a reduction in ranking error, but the paper does not report specific values. According to Table 2, the Oracle@64 for Stage 2 decreased from 0.9976 to 0.9939, a drop of 0.37 percentage points; the gap narrowed from 0.0563 to 0.0491, a reduction of 0.72 percentage points. When these two figures are offset against each other, the result is exactly a 0.35 percentage point improvement in top-1. Please provide this breakdown and explain exactly how Stage 2 improved the candidate distribution itself.
-
Does the “frozen teacher” in Stage 2 replicate a copy from the current student in each cycle, or does it remain unchanged throughout the entire 30 cycles? If the former, it effectively acts as a proximal constraint on the previous round, and the entire Stage 2 amounts to repeatedly performing filtered-BC; if the latter, the target set is already fixed, so I don’t understand why 60 epochs are still needed.
-
What exactly is the configuration for “Non-iterative generator/scorer update” in Table 8? Does it mean the generator and scorer are backpropagated together in the same batch, or does it stop after just one round of scorer updates and one round of generator updates? What kind of trajectory does the model output when it crashes below 10 points?
-
The range of PDMS for the three seeds in Table 21 is only 0.014, and the EPDMS is also 0.014 which is unusually small. Did these three runs use the same Stage 1 checkpoint? Please specify exactly what changes were made between the three runs.
7 comments
Q4. Does the Pareto Set Degenerate into Selecting the Highest-EP Trajectories?
We believe the distinction between maximizing EP and selecting high-quality trajectories under multiple planning objectives is important.
First, the performance differences are not restricted to EP. In the NAVSIM v1 comparison (Table 1), CLOVER ranks first or second across all reported sub-metrics. On NAVSIM v2 navtest and NavHard, CLOVER does not achieve the highest EP among the compared methods, yet remains highly competitive in overall planning performance.
More generally, an appropriate autonomous driving objective should encourage progress toward the destination while satisfying safety, traffic-rule, and comfort requirements, rather than maximizing progress in isolation.
Regarding the Pareto set, its construction provides a direct mechanism for preserving these competing objectives.
From a multi-objective optimization perspective, Pareto selection operates on a vector of planning metrics rather than a single scalar score. A trajectory with lower EP can remain non-dominated if it provides improvements in other objectives that cannot be simultaneously matched by a higher-EP alternative.
Therefore, unless the different planning objectives admit no meaningful trade-offs, Pareto selection does not reduce to simply choosing the trajectories with the highest EP. Its purpose is precisely to preserve meaningful trade-offs that would otherwise be discarded by a single-objective selection rule.
Furthermore, NAVSIM's EP is normalized against a reference progress value and capped at 1. Once this upper bound is reached, traveling farther does not yield an additional EP reward.
A representative example is scene a1eab14904fc5807. Among the 64 candidate trajectories, we identified the following two non-dominated solutions under the true evaluator:
| Candidate | NC | DAC | EP | TTC | Comfort | DDC | PDMS |
|---|---|---|---|---|---|---|---|
| #25 | 1 | 1 | 1.0000 | 0 | 0 | 0 | 0.4167 |
| #44 | 1 | 1 | 0.9496 | 1 | 1 | 1 | 0.9790 |
Candidate #25 achieves the maximum EP of 1.0, but performs poorly on TTC, comfort, and driving-direction compliance.
Candidate #44 sacrifices approximately 0.05 EP while substantially improving the other metrics, increasing its final PDMS from 0.4167 to 0.9790.
Both trajectories lie on the true Pareto frontier because neither dominates the other across all objectives. A selection rule based solely on maximum EP would discard candidate #44, despite its much higher overall planning quality.
This example directly illustrates why multi-objective Pareto supervision provides information beyond EP maximization.
Therefore, our results do not support the hypothesis that Pareto selection has degenerated into selecting only the highest-EP trajectories.
Regarding the relatively low EC without an anchor, we attribute this primarily to the interaction between independent cross-frame selection and the pseudo-closed-loop evaluation protocol, as discussed further in W1.
Q5. Numerical Decomposition of Eq. 97 and the Actual Contribution of Stage 2
We agree with the numerical decomposition provided by the reviewer.
According to Eq. 97:
From Stage 1 to Stage 2:
- Oracle@64: 0.9976 → 0.9939, a change of −0.0037.
- Ranking gap: 0.0563 → 0.0491, a reduction of 0.0072.
- Selected Top-1: 0.9413 → 0.9448, an improvement of 0.0035.
Thus,
This confirms that the improvement in the selected Top-1 trajectory is primarily reflected in a reduction in ranking regret, rather than a further increase in the already near-saturated Oracle@64.
However, we would like to distinguish this exact statistical decomposition from the independent contributions of the generator and scorer.
Because Stage 2 alternately optimizes both components, the proposal distribution and the scorer evolve together. The oracle score and ranking gap are therefore measured on different proposal distributions at different stages, making it difficult to attribute their changes to either component independently.
To investigate this interaction, we conducted an additional experiment by exchanging the generator and scorer checkpoints between Stage 1 and Stage 2:
| Generator | Stage 1 Scorer | Stage 2 Scorer |
|---|---|---|
| Stage 1 Generator | 94.13 | 93.6 |
| Stage 2 Generator | 93.3 | 94.5 |
Neither replacing the generator alone nor replacing the scorer alone reproduces the final Stage-2 performance.
Using the Stage-2 generator with the Stage-1 scorer yields 93.3 PDMS, while using the Stage-1 generator with the Stage-2 scorer yields 93.6 PDMS.
Only the jointly refined Stage-2 generator and scorer achieve 94.5 PDMS.
This indicates that the central benefit of Stage 2 lies in the co-adaptation of the generator and scorer to the evolving proposal distribution, rather than an isolated improvement in either module.
Stage 2 also substantially improves the overall quality of the candidate distribution:
- Mean@64: increases from 0.7374 to 0.8277.
- Number of low-quality candidates (PDMS below 0.50): decreases from 12.21 to 6.83 per scene.
- Pairwise ADE: remains at 5.20, substantially higher than the baseline value of 1.80.
These results clarify the complementary roles of the two stages.
Stage 1 primarily expands proposal coverage and improves the oracle upper bound. Stage 2 then refines this expanded distribution by removing low-quality candidates, improving overall proposal quality, and enabling the scorer to select better trajectories from the updated candidate set.
Q6. Frozen Teacher Updates and the Meaning of 60 Epochs
The former interpretation is correct. At the beginning of each alternating cycle in Stage 2, we create a frozen teacher by copying the current student generator.
During the generator-refinement phase, the teacher remains frozen and provides the top-k and Pareto distillation targets. The stability regularization term constrains the student's updates relative to the frozen teacher.
At the next cycle, the teacher is refreshed from the updated student. The scorer is also refitted on the current proposal distribution, producing updated guidance for subsequent generator refinement.
Therefore, the reviewer's interpretation of the frozen teacher as a proximal constraint relative to the preceding iteration is reasonable. Its purpose is to restrict the update magnitude and prevent excessive exploitation of scorer estimation errors.
Importantly, the teacher proposals, scorer, and selected targets evolve across cycles. Stage 2 is therefore not repeatedly fitting a single fixed set of teacher trajectories.
The training schedule consists of 30 alternating cycles, each containing:
- One epoch of scorer fitting.
- One epoch of generator refinement.
This corresponds to 30 × 2 = 60 training epochs.
The repeated scorer-fitting phases are necessary because generator updates continuously change the candidate distribution on which the scorer is expected to operate.
We thank the reviewer for the detailed analysis of our method and experiments. The questions raise important points regarding anchor configurations, Pareto supervision, the contributions of the two training stages, and the mechanisms underlying alternating optimization. We address the eight questions first, followed by the two major concerns raised in the Critical Weaknesses section.
Responses to Questions
Q1. Dataset Split Used for the Parameter Search in Table 13
We followed the
train_logs/val_logsconfiguration provided in the NAVSIM training setup.Specifically, the
val_logssubset belongs to the NAVSIM navtrain data domain, rather than the independentnavtestevaluation split. Since this subset was also included in model training, it is more accurately described as an internal subset used for parameter selection rather than a strictly independent held-out validation set.This setup is consistent with the training configurations commonly adopted in NAVSIM-based planning methods.
We will clarify the dataset partition and parameter-search procedure in the revised manuscript.
Q2. Anchor Usage and Results Without Anchors on NavHard
Yes, the NavHard results in Table 4 use an anchor. The main results reported under the EPDMS metric use the corresponding anchor-assisted configuration, which was not sufficiently explicit in the original presentation.
We additionally evaluated CLOVER on NavHard without an anchor:
The results follow the same trend as the NAVSIM v2 navtest evaluation.
Without an anchor, CLOVER achieves higher EP in both stages, while EC decreases substantially. The anchor provides a more favorable trade-off between trajectory flexibility and temporal consistency.
We will explicitly label the anchor configurations in the revised tables.
Q3. Inference Latency with the Anchor
The reported 110 ms includes only the primary CLOVER model.
The default anchor uses the same base architecture with a separate set of weights. Since the two models perform largely independent inference, they can be executed concurrently.
The measured total inference latency with the default anchor is 137.2 ms on a single NVIDIA A100 GPU, compared with 110 ms for the primary CLOVER model alone.
This is substantially faster than fully sequential execution of the two models, which would take approximately 220 ms.
We will clarify this inference configuration and its latency in the revised paper.