acceptodds
Under review as a conference paper at ICLR 2027

Vault: One-Step Latent Generation with Positive-Anchored Rewards for Autonomous Driving

Abstract

End-to-end autonomous driving must balance multi-modal maneuver generation against real-time inference constraints. Diffusion planners capture diverse behaviors but their iterative denoising is too slow for safety-critical deployment, while one-step alternatives still imitate only the single expert demonstration, leaving neither diversity nor quality certified. To address this, we propose Vault, a framework that couples one-step latent generation with sample-based reinforcement learning guidance. Vault generates candidates in a single forward pass by drifting in a VAE latent space, built on a V-JEPA 2.1 visual representation pretrained on the full NAVSIM training imagery. During training, a reward-gated positive pool collects the full-score trajectories that the model itself discovers under the official evaluation and anchors the drift targets to these certified samples. This closes a reinforcement-learning loop purely through sample-based target construction, without policy gradients, learned reward models, or extra interaction: the generator imitates the frontier of its own certified successes, so quality and multi-modality improve together under a single fixed sampling configuration. A learned scorer predicting the official score and its sub-metrics selects the executed trajectory, and all training-only modules are discarded at inference. On the NAVSIM benchmarks, Vault achieves state-of-the-art closed-loop performance in real time.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Autonomous Driving, Trajectory Planning, One-Step Generation, Drift-Based Generative Models

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

This paper proposes Vault, a framework for autonomous driving planning. It replaces diffusion planner to one-step latent drifting, then repeatedly collects maximum-scoring self-generated trajectories under official evaluator as imitation goals for efficient planning and maintaining multimodality. The proposed method achieves state-of-the-art performance on both NAVSIM v1 and v2.

Critical strengths
  1. The proposed method achieves both SOTA PDMS scores on both NAVSIM v1 and v2 compared with substantial amount of recent baselines.
  2. The proposed Vault framework's design utilize the efficiency of one-step generation to avoid the iterative denoising procedure of Diffusion planner via Drifting, also with a self-imitation mechanism to expand imitation target pool using high-quality self-generated trajectories.
Critical weaknesses
  1. The Vault use the V-JEPA 2.1 pretrained visual backbone and other competitive baselines use different backbones, which make it difficult to attribute the improvements from proposed planning modules and backbone. Given that Valut only surpass the best baseline by 0.1 and 0.2 PDMS score on v1 PDMS and v2 EPDMS respectively, a more controlled comparison and ablation on both v1 and v2 could further strengthen the proposed method.
  2. The presentation of the method could improve for clarity. Vault integrated several components from previous work, but the motivation for combining these components and their individual roles are not always very clear, especially in the introduction. A clearer presentation of the overall method and the motivation behind each component would make the methodological contribution easier to understand.
Questions for authors
  1. The authors could add more related citations like End-to-End Driving with Online Trajectory Evaluation via BEV World Model (ICCV25), Driving on Registers: Simple and efficient transformer based end-to-end driving (CVPR 2026).
  2. Vault achieves notable higher EP on both v1 and v2 compared with most baselines. The paper attributes part of this improvement to the reward-gated positive pool providing higher-progress imitation targets. Could author provide more analysis or intuition on why Vault particularly improves on EP instead of TTC or other sub metrics?
  3. The paper could provide more analysis on Vault design to better explain why works. For example, it is helpful to understand the method by showing how the positive imitation pool evolves during training, visualization on distributions of the positive and negative points in latent space, and more detailed comparison of multimodality planning results with original diffusion planner.
3 comments
AuthorsOpenReview ↗

Controlled backbone test, clarified design motivation, and mechanism analyses

We thank the reviewer for the careful summary and suggestions. The revision adds a controlled backbone experiment, a matched-budget pool control, a NavHard two-stage evaluation, and a quantitative diversity appendix, and rewrites the introduction. All changes are marked in blue.

CW1: We agree this needed a controlled test and have added one (Sec. 4.2, blue). Replacing the visual stack with a LoRA-fine-tuned DINOv2-S at 756×1344 (the backbone of DrivoR and CLOVER), with the planning pipeline otherwise unchanged, reaches 94.5 PDMS with EP 92.8: it matches CLOVER, exceeds DrivoR by 0.8, and sits only 0.1 below the V-JEPA variant, within the five-seed spread (94.52–94.60). The visual stack thus contributes about 0.1, and the planning design carries the performance. The ego-progress margin over CLOVER (92.8/92.9 vs. 91.7) holds on both backbones. All other controls likewise hold the backbone fixed. The controls are on v1, where the strongest baselines report. On v2, Vault's corrected 91.2 exceeds SparseDriveV2 (90.1), CLOVER (90.4), and SUV (91.0) with the highest mean EP (92.1), and on NavHard Vault leads DrivoR and CLOVER by 1.0 (49.3 vs. 48.3) with the highest stage-1 EP (79.5).

CW2: Vault rests on two coupled ideas. First, existing generative planners run or compress the denoising chain at inference. Drifting instead learns the scene-conditional distribution directly during training, so a single forward pass yields multimodal samples. Second, this is what makes reward guidance natural: drifting's supervision is an anchor set, so the reward acts simply by deciding which samples become anchors. The pool stores every self-discovered perfect-score trajectory and lets supervision evolve with the policy, with no policy gradients, learned reward models, or extra interaction. The remaining components are deliberate adoptions: the pipeline follows the mainstream generate-and-select paradigm, and the backbone follows Drive-JEPA in adopting a V-JEPA, since a predictive representation of the image stream suits an anticipatory task. The VAE/PCA drift spaces are engineering instantiations (Sec. 3.2, App. B–C). The introduction now states this causal chain explicitly.

Q1: WoTE is now cited in Related Work (blue): it evaluates candidates online through BEV world-model rollouts, whereas Vault's scorer predicts the official metric directly from the current observation, without simulation. DrivoR was already cited in six places. We use its no-SimScale variant for comparability.

Q2: The perfect-score gate explains the pattern, in two steps. First, the maximum score admits any collision-free trajectory that advances at least as far as the reference, so the pool can contain and the generator can learn farther-advancing trajectories. This matters because the human reference itself scores only EP 87.5 on navtest (the simulated rollout travels less far than planned), so longer certified anchors directly raise the progress the planner can express. Second, the supervision is nearly flat along safety dimensions: among feasible candidates, satisfying versus violating TTC differs only marginally, and NC, DAC, TTC are near their ceilings (99.0/98.8/96.4), while progress varies widely among perfect-score trajectories. Comfort is honestly traded (99.0 vs. 99.9 human). Three results support this:

  • The pool lifts EP 91.8→92.9 with total 93.9→94.6 (Table 1).
  • The feasibility-only control, with the same budget and refresh but admission by feasibility regardless of score, stays at 93.9 with EP 92.0, recovering none of the gain (Sec. 4.4, blue).
  • On NavHard the gain again concentrates in ego progress (stage-1 EP 79.5, the highest among compared planners).

Q3: We added a quantitative diversity analysis (App. D, blue). With the pool, candidate sets are roughly five times wider (pairwise ADE 1.95 vs. 0.38 m) and span 10.75 certified endpoint clusters (Q-Cluster@2 m over candidates with true PDMS ≥ 0.8) versus 1.71, and the oracle ceiling rises from 98.0 to 99.3. The expert-only model in fact has more high-scoring candidates (895.7 vs. 774.5; mean score 0.853 vs. 0.792): the pool contributes certified mode coverage, not candidate count or average quality, which is why quality and multimodality improve together. Against iterative denoising at the same backbone, scorer, and budget (Sec. 4.3), the diffusion decoder trades progress for comfort (92.0 vs. 94.6; EP 86.5 vs. 92.9; comfort 100.0 vs. 99.0). On pool evolution: the pool refreshes once per epoch (App. G) and per-scene anchor coverage saturates within the first 2–3 epochs, reaching the steady-state branches of Fig. 2 (later epochs only add density), and since each iteration subsamples at most M = 32 anchors, the effective supervision stabilizes early. App. C (Fig. 5, Table 4) already visualizes the VAE latent space against physical and PCA spaces. We are happy to post per-epoch pool statistics or latent anchor plots during the discussion period.

Thank you to the authors for their detailed response and additional experiments, particularly on NavHard and the controlled-backbone ablation. However, I remain concerned about the limited improvements on NAVSIM v1 and v2 and the missing v2 ablations. In particular, the new Vault variant achieves the same PDMS as the strongest baseline. Given the remaining concerns, I find it difficult to justify raising my score. Therefore, I will maintain my score.

AuthorsOpenReview ↗

Margins against the published baselines, consistency across v1, v2, and NavHard, and the cost of v2 ablations

Thank you for re-reading our response and for stating the remaining concerns so plainly. We would like to add one observation about the margins before turning to the two open points.

The 0.1 and 0.2 gaps mentioned in the review are gaps to the two newest preprints, CLOVER and SUV. Against the baselines that have been through peer review, the margins are not close. DrivoR (CVPR 2026) trails by 0.9 on v1 navtest (93.7 vs. 94.6), by 0.8 under its own backbone family (93.7 vs. 94.5 for our DINOv2-S control), and by 1.0 on NavHard (48.3 vs. 49.3). DriveSuprim (AAAI 2026) trails by 1.1 on v1 (93.5 vs. 94.6). In Table 3 it reports only under the original evaluator, where it trails by 2.3 (86.0 vs. 88.3). The compressed top of the table therefore involves only the two most recent entries, and we stay ahead of both on every split they report.

The pattern we care more about is consistency. Vault leads v1 (94.6), v2 navtest (91.2 corrected, with the highest mean ego progress at 92.1), and NavHard (49.3, against 48.3 for DrivoR and CLOVER). The closest competitors each fall away somewhere. CLOVER, the strongest overall, gives back a full point on NavHard. SUV, closest on v2 at 91.0, runs v2-only experiments in its main paper. The v1 number in its supplementary is 90.8, which is 3.8 behind ours. DrivoR is behind on every split we share. On v1 the top method scores sit within 0.3 of the human reference (94.8), so the headroom there is thin. Ego progress is where the methods still separate, and Vault leads it on both versions (92.9 and 92.1).

On the DINOv2-S variant, we added this run during the discussion period in response to CW1. The swap is deliberately plain. We replace the visual encoder with a LoRA-fine-tuned DINOv2-S at 756×1344 input and change nothing else in the pipeline. There is no DINO-specific tuning, and we did not adjust the decoder, the scorer, or any capacity to accommodate the new encoder. The variant still reaches 94.5 under the baselines' own backbone family, on par with CLOVER and 0.8 above DrivoR, with ego progress again higher (92.8 vs. 91.7). Losing only 0.1 to the default model after a drop-in swap tells us two things. The gain sits in the planning design, not the backbone. And the pipeline takes to a new visual stack easily, which we take as a practical strength of the method.

On simplicity, the deployment story is identical across the three systems. One forward pass generates the candidates and a scorer picks one. The difference is the training wrapped around it. DrivoR keeps its generator on plain imitation and lets evaluator feedback reach the scorer only. CLOVER needs more machinery. It builds pseudo-expert trajectories from privileged metric-cache information (route centerlines, drivable-area maps, future occupancy), then runs thirty cycles of alternating scorer fitting and generator self-distillation. For v2 it further adds a separately trained anchor model with a tuned soft reranking, which supplies 1.1 of its reported 90.4 (89.3 without it). Vault trains the whole planner in one loop. The evaluator only decides which self-generated samples become anchors for the drift loss. The scorer never sits in the generator's training path, and nothing is attached at inference. Our 91.2 on v2 carries no reranking. Reaching these numbers with less machinery is a fair part of what the paper argues for.

On the missing v2 ablations, the v1 and v2 planners are separate runs. Each run has its own evaluator-gated pool, and its scorer learns from that evaluator's labels. EPDMS is also a different objective, adding DDC, TLC, LK, HC, and EC on top of the v1 terms. An ablation cell on v2 is therefore a full second training campaign, not a re-evaluation. Within the discussion window we put the controls on v1, where the mechanism was designed. NavHard served as a second, harder split, and the gains persist there. This is also where the published baselines place their ablations. CLOVER's component ablations (its Tables 6–8) are all on v1, with only the v2-specific reranking ablated on v2. DrivoR goes further and computes all ablation scores on v1 navval, not even on the test split.

We hope the consistency across the three settings, together with the attribution controls, makes both the size and the origin of the improvements easier to judge.

Reviewer_xkk5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Vault, a one-step latent trajectory generation framework for end-to-end autonomous driving. The model generates a set of trajectory candidates in one forward pass by applying a drifting objective in a VAE latent space. Its main novelty is a reward-gated positive pool: during training, self-generated trajectories that achieve the maximum score under the official evaluator are stored and subsequently used as positive anchors together with the logged expert trajectory. A learned scorer then predicts the official score and its sub-metrics to select the final trajectory. On NAVSIM v1 and v2, Vault reports 94.6 PDMS and 91.2 corrected EPDMS, respectively, with an end-to-end latency of 91.4 ms for 1024 candidates.

Critical strengths
  1. From my perspective, the reward-gated positive pool is an interesting idea and the major novelty of the paper. It provides a simple way to combine one-step drifting-based generation with reward-guided self-imitation. Compared with learning from one single demonstration, the pool can expose the generator to multiple high-scoring candidates for the same scene.

  2. The method achieves strong results on both NAVSIM v1 and v2. In particular, the positive pool mechanism improves v1 PDMS from 93.9 to 94.6, with the largest gain coming from ego progress, which is consistent with the motivation of collecting successful trajectories that can make more progress than the expert demonstration.

  3. The one-step generator is computationally efficient. Although the complete system still evaluates 1024 proposals, the decoder itself takes only 5.4 ms, and the reported end-to-end latency satisfies a 10 Hz planning frequency on an RTX 4090.

Critical weaknesses
  1. The experimental validation is mainly limited to the Navtest split of NAVSIM. While NAVSIM provides useful non-reactive trajectory evaluation, it cannot fully evaluate how the planner behaves when its actions affect subsequent observations and other agents. Evaluating Vault on a closed-loop benchmark such as HUGSIM [1], or at least on a more challenging open-loop benchmark such as Navhard, would strengthen the empirical results of the paper.

  2. The ablation study is somewhat insufficient and is conducted only on NAVSIM v1. A few detailed design choices have not been verified, such as the pool size and refresh frequency. Line 404 mentions that "A paired pool-versus-expert comparison and a pool-size sweep would provide a more direct test of this mechanism". However, I could not find such experiments in the paper, which makes this statement a bit confusing.

  3. The generator comparison does not cleanly isolate the benefit of one-step drifting. As the paper acknowledges in Section 4.3, replacing Vault with a diffusion decoder changes both the decoder architecture and the training objective. Therefore, the 2.6-point difference cannot be attributed specifically to one-step generation or to the drifting objective. I would recommend to add a more controlled comparison between diffusion and drifting models.

  4. As a minor concern, I do not find the oracle best-of-N results in Table 1 particularly informative, as prior work has shown that a near-perfect PDMS can be achieved with a sufficiently large number of candidate trajectories [2].

[1] Zhou H, Lin L, Wang J, et al. Hugsim: A real-time, photo-realistic and closed-loop simulator for autonomous driving[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. [2] Yao W, Li Z, Lan S, et al. Drivesuprim: Towards precise trajectory selection for end-to-end planning[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2026, 40(14): 11910-11918.

Questions for authors

I don't have other questions. Please refer to the Critical Weaknesses section.

2 comments
AuthorsOpenReview ↗

New NavHard results, matched-budget pool control, and diversity analysis

We thank the reviewer for the careful reading and for identifying the reward-gated positive pool as the central contribution, which is indeed the mechanism we set out to isolate. The revision adds a NavHard two-stage evaluation, a matched-budget feasibility-only pool control, and a quantitative diversity appendix, and rewrites the sentence quoted in CW2; all changes are marked in blue.

CW1: We have added the NavHard two-stage evaluation (Sec. 4.2, Table 2, blue). Its follow-up scenes branch from the planner's first-stage outcome, so first-stage errors propagate into the second stage, approximating closed-loop behavior within NAVSIM. Vault reaches 49.3 combined EPDMS, above DrivoR and CLOVER (both 48.3) and GTRS-D (45.0) in the same no-simulated-augmentation setting, with the highest stage-1 ego progress (79.5). A full closed-loop benchmark such as HUGSIM is the natural next step, but the discussion period is too short for us to train and tune for a full closed-loop setting, so we leave it as future work; Vault already meets the real-time prerequisite (91.4 ms end-to-end, 10 Hz on an RTX 4090).

CW2: The quoted sentence was indeed a leftover promise of experiments that the paper did not contain; we apologize and have rewritten it, with the replacement backed by a new experiment. The feasibility-only control (Sec. 4.4, Table 1, blue) is the matched pool-versus-expert comparison: it keeps the expert anchor, the anchor budget, and the once-per-epoch refresh, and changes only the admission rule (drivable and collision-free, regardless of evaluator score). It scores 93.9 with EP 92.0, on par with the expert-only variant and recovering none of the 0.7-point pool gain, so the maximum-score certification, rather than the anchor count or the refresh, raises the supervision frontier. A pool-size sweep is now explicitly stated as future work. On the v1-only scope: the v2 planner is a separate training run with its own evaluator-gated pool (App. G), so each v2 control costs a full run; we therefore ran the mechanism controls on v1, directly comparable to the baselines, and added NavHard as a second, harder split on which the gains persist (the v2 results, corrected 91.2 with the highest mean EP of 92.1, are consistent with the same mechanism).

CW3: We agree, and the revision keeps the explicit disclaimer: Sec. 4.3 states that the swap changes the decoder and the training objective, "so it does not by itself isolate the drift objective." We report it as a system-level comparison at matched backbone, scorer, training scenes, and budget (92.0 vs. 94.6, with the diffusion variant trading ego progress for comfort: EP 86.5 vs. 92.9, comfort 100.0 vs. 99.0), and do not attribute the 2.6 points to the drift objective alone. Table 1 also contains a more controlled contrast: Vault without positive pool and Vault with diffusion decoder share the same logged-expert supervision, backbone, scorer, and N = 1024 budget, so neither uses multi-anchor supervision, and the one-step drifting generator still reaches 93.9 versus 92.0. The remaining attributions are component-wise: the pool ablation isolates the pool (93.9 to 94.6, plus the feasibility-only control above), and the DINOv2-S variant isolates the visual stack (94.5 vs. 94.6, within the five-seed spread). The clean test the reviewer describes, in which only the diffusion objective is swapped within the same VAE latent space, is less straightforward than an objective swap, and we agree that running it as future work would strengthen the attribution: the diffusion decoder regresses toward a single target per scene (the logged demonstration); replacing it with one pool positive is feasible in code, but with the set of certified positives it is not, since denoising regression has no per-scene multi-target supervision, and accommodating an anchor set is precisely the structural difference the drift objective introduces.

CW4: The oracle rows are not headline results; they carry two specific analyses. Mainly, they verify that the drift space does not compromise generation completeness: the generator operates through a compressed pipeline (a VAE latent space with a frozen PCA basis) and under oracle selection still reaches 99.3 PDMS at N=1024, so essentially every scene admits near-perfect trajectories the model can produce; neither component renders trajectories ungeneratable, and the residual 4.7-point gap to the deployed score is a selection problem, not a generation deficiency. Second, comparing the full and no-pool oracles at each N (98.4/99.0/99.3 vs. 97.2/97.8/98.0) isolates the pool's effect on the generation ceiling. We agree, and with DriveSuprim that we already cite, that a sufficiently large N approaches near-perfect PDMS; this is why the rows are reported at three N values as ceiling comparisons, and why the diversity appendix measures certified mode coverage at fixed N instead.

We hope the new results address the empirical concerns.

Thanks for the reply. I find the new NavHard results to be promising and I will keep my current score.

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

this work proposes a one-step latent generation method for autonomous driving that integrates three techniques: one-step latent drifting, reward-gated positive anchors, and trajectory ranking. In terms of performance, Vault achieves 94.6 PDMS (v1) and 91.2 EPDMS (v2) on the NAVSIM benchmark. The reviewer finds the integration useful but notes that its components largely build on existing Drifting, self-imitation, and off-the-shelf scorers, and that the distinctive contribution and claimed diversity gains are not sufficiently isolated.

Critical strengths

1/The experimental coverage is reasonably broad, including two NAVSIM benchmarks, a positive-pool ablation, an oracle analysis, and latency measurements. 2/The paper is generally readable, with clear organization and useful tables.

Critical weaknesses

1/The paper states that "V has a batch RMS field magnitude and a scalar value of exactly one", it should specify whether V covers one candidate or N candidates, since the normalization and constant-loss derivation depend on it. 2/The paper states that "it uses a squared-distance kernel in Eq. (1) and distance-based bidirectional weights in Eq. (3)", it should explain how the modified kernel/normalization relates to the reference field, because its theory does not transfer automatically. 3/The paper states that "it uses scale floors epsilon/epsilon_s and a ranking margin delta", the values of epsilon, epsilon_s, and delta should be reported to make the loss and scorer reproducible. 4/ "Eq. equation 1/5" should use "Eq. (1)/(5)" consistently. 5/ "d(z_i, a_j) and average d(z, A)", the distance, averaging indices, and normalization must be defined at first use.

Questions for authors

see the above.

1 comment
AuthorsOpenReview ↗

All five points fixed, new backbone, pool, and diversity isolation analyses

We thank the reviewer for the close reading. All five points are fixed in the revised PDF, with changes marked in blue; none of these five fixes changes a claim or a reported result. We answer each point and then the broader concern raised in the summary.

CW1: λ is a single scalar per training step, shared by all B×N candidates of the batch; it is neither per-candidate nor per-scene. The subscript i in the submission was a notational slip on our part, now removed. The overline is defined at first use (Sec. 3.2) as the batch RMS, and App. B.4 now states that the floor enters additively inside the square root, so the loss equals v/(v+ε), with v the batch mean squared field magnitude; that is, one up to the floor term. Eq. (5) is written per scene for readability; the implementation batch-averages under the same λ. This batch-level granularity is also the reference implementation's own choice (the drift normalization of Deng et al., whose expectation is likewise replaced by the batch mean).

CW2: In Eq. (1) (line 168 of the original submission), the subscript 2 in k(x, y) = exp(−‖x−y‖₂/τ) denotes the ℓ₂ norm, which matches both the Euclidean distance d(z_i, a_j) = ‖z_i − a_j‖₂ in Eq. (3) and Deng et al.'s main kernel in their Eq. (12); the squared form appears in their paper only in the MMD discussion (their App. C.2), so the two forms are easy to conflate. Moreover, several components in Eqs. (3)–(5) follow their reference implementation directly: the scale S_sc corresponds to their feature normalization (their App. A.6, Eq. (21): batch-empirical, stop-gradient, √C-scaled mean distance), the RMS step size λ is their drift normalization (their Eq. (25)), the multi-bandwidth sum follows their multi-temperature aggregation with τ̃ = τ√C (their Eq. (22)), which matches our effective bandwidths {√D, 2√D, 4√D}, and the bidirectional √(r_ij c_ij) weights follow their Alg. 2. The departures that are genuinely ours are three: we normalize the anchor and negative blocks separately, we mix conditional and unconditional anchors through α, and we replace fixed positives with the reward-gated pool. For these we agree the reference guarantees do not transfer automatically. The theory we rely on is the fixed-point condition of Eq. (9), which is derived for the modified field rather than inherited; the revision adds these correspondences to App. B.1/B.2/B.4.

CW3: The values are now reported in App. G: ε = 10⁻⁶ (additive inside the square root, so λ ≥ 10⁻³), ε_s = 10⁻⁴, and δ = 0.02 on the [0,1] PDMS scale. λ_rank and λ_mse were already given in App. E.

CW4: This was a packaging error, our apologies: a documentation macro silently overrode amsmath's \eqref, so every reference rendered as "Eq. equation N". Fixed.

CW5: d(z,z′) = ‖z−z′‖₂ and the overline average, over all candidate-anchor pairs of the batch, are now defined at first use in Sec. 3.2; App. B.1 is consistent.

Summary: The revision adds the isolation the summary asks for, all under a fixed backbone and scorer.

  • A DINOv2-S variant matches CLOVER at 94.5 with the V-JEPA model only 0.1 higher, so the visual stack is not where the margin comes from.
  • The diffusion-decoder swap costs 2.6 PDMS, removing the pool costs 0.7, and a new matched-budget control (same anchor budget and refresh, admission by feasibility only) recovers none of that 0.7, which points to the score gate itself rather than the number of anchors.
  • A new appendix quantifies diversity: the expert-only model collapses to a single mode (pairwise ADE 0.38 m, 1.71 certified clusters among 895.7 candidates with PDMS ≥ 0.8), while the full pool spans 10.75 clusters at five times the spread; the pool's contribution is certified mode coverage, not candidate count.

We hope this clarifies both the notation and the contribution structure.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions: four new controlled experiments and clarifications

We thank all reviewers for their constructive feedback. In the revised manuscript (changes marked in blue), we have added four new controlled experiments/analyses and clarified all mathematical and implementation details:

  1. Controlled visual-backbone ablation (Sec. 4.2, Table 1; R1, R2). Replacing the V-JEPA 2.1 backbone with a LoRA-fine-tuned DINOv2-S, the backbone used by DrivoR and CLOVER, achieves 94.5 PDMS with EP 92.8, within 0.1 of the full model (94.6) and matching CLOVER with a +1.1 EP margin, and the planning design, not the visual stack, carries the performance.

  2. Matched-budget feasibility-only pool control (Sec. 4.4, Table 1; R1, R2, R3). Admitting self-generated trajectories by feasibility alone (collision-free and drivable, without maximum-score certification) under the same anchor budget and refresh schedule yields 93.9 PDMS (EP 92.0), recovering none of the +0.7 gain of the full reward-gated pool (94.6), which isolates the score gate as the active ingredient.

  3. NavHard two-stage evaluation (Sec. 4.2, Table 2; R2, R3). On this pseudo-closed-loop split, Vault reaches 49.3 combined EPDMS, above DrivoR and CLOVER (both 48.3) and GTRS-D (45.0) in the same no-simulated-augmentation setting, with the highest stage-1 ego progress (79.5).

  4. Quantitative diversity analysis (App. D, Table 5; R1, R2, R3). Following the protocol of CLOVER, the positive pool expands certified mode coverage from 1.71 to 10.75 endpoint clusters (among candidates with PDMS ≥ 0.8), widens pairwise ADE by 5× (0.38 m → 1.95 m), and lifts the oracle ceiling from 98.0 to 99.3.

  5. Clarity and reproducibility (Sec. 1, Sec. 3.2, App. B, App. G; R1, R2). The introduction is rewritten to state why one-step drifting and the reward-gated pool belong together. The correspondences with Deng et al. (2026) are detailed in App. B. All numerical floors and margins (ε, ε_s, δ) are reported in App. G.

Detailed point-by-point responses are posted as replies to each review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.