acceptodds
Under review as a conference paper at ICLR 2027

ReTranDrive: A Reinforced Transition-Aware Framework for Vision-Language-Action Autonomous Driving

Abstract

Autonomous driving policies are commonly optimized by evaluating the quality of each predicted plan in isolation, despite being executed through continuous replanning. This formulation overlooks a distinct aspect of sequential decision-making: the quality of transitions between consecutive plans. As a result, a planner can produce individually high-quality trajectories while exhibiting unstable behavior over time, repeatedly altering its intended motion without a corresponding change in the scene. The missing ingredient is an explicit objective over inter-plan transitions, capturing whether successive predictions evolve coherently with the changing scene rather than merely optimizing each plan in isolation. We present ReTranDrive, which defines this objective by shifting the optimization target from single-step planning quality to sequential execution coherence: scoring a plan against the current observation alone becomes a transition-aware formulation over consecutive plan pairs. Applied to a VLM-conditioned diffusion planner as a differentiable term during imitation and a simulator-verifiable reward during reinforcement, it regularizes only those inter-plan changes not warranted by scene evolution rather than enforcing unconditional temporal consistency. To resolve the credit-assignment challenge inherent in pairwise scoring, we introduce an asymmetric scheme that assigns inconsistency strictly to the decision responsible for it. On the NAVSIM benchmark, ReTranDrive attains a state-of-the-art EPDMS of 90.94 under the extended metric and a PDMS of 94.40, reaching 99.58% of human driving performance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Autonomous Driving, Vision-Language-Action (VLA) Models, Diffusion Policy, Reinforcement Learning, Sequential Decision-Making

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZaHtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ReTranDrive addresses inconsistency between successive plans during repeated replanning. It augments a VLM conditioned diffusion planner with a differentiable kinematic consistency loss during imitation learning and an extended comfort reward during reinforcement learning. During reinforcement, a frozen Stage 2 model provides the reference plan from the previous observation, while updates apply only to the current prediction. Experiments on NAVSIM navtest report 94.40 PDMS and 90.94 EPDMS, with ablations comparing training stages, reward forms and reference choices.

Critical strengths
  1. The paper addresses a practically relevant limitation of optimizing each plan in isolation. Its consistency objective is simple and requires no additional learned module for coherence. Treating consistency as a soft objective allows it to be balanced against safety and progress.
  2. Table 4 provides direct evidence of improvement over the same planner using a reward based only on the current plan. EPDMS increases from 90.45 to 90.94 and EC from 65.13 to 69.80, while EP remains nearly unchanged at 93.39 versus 93.40. The full system also achieves strong overall benchmark scores.
Critical weaknesses
  1. Novelty is limited, and the claimed selective consistency is insufficiently demonstrated. The method adapts the existing EC metric within an established VLM diffusion planning and group relative optimization framework. EC applies fixed thresholds to kinematic differences without directly assessing whether scene changes require replanning. No targeted experiment separates necessary from unnecessary revisions. A simple temporal smoothing baseline and an isolated ablation of the Stage 2 consistency loss are also missing.
  2. The results do not fully establish improved coherence during continuous execution. EC is 69.8, below most methods in Table 2 despite strong aggregate scores. Although Table 5 compares reference choices, the main training objective uses a frozen Stage 2 reference rather than the deployed policy’s own previous plan. Evaluation is limited to the paper’s nonreactive NAVSIM setting, leaving benefits under repeated execution and responsive traffic interactions unverified.
Questions for authors
  1. Can you separately evaluate scenes requiring little plan revision and scenes requiring a timely change? Reporting reaction delay and safety outcomes would help determine whether improved consistency preserves necessary responses.
  2. Can you isolate the Stage 2 consistency loss while keeping the diffusion expert unchanged, and compare against simple temporal smoothing under the same architecture and training budget?
  3. Beyond the reference comparison in Table 5, what evidence shows that consistency with a frozen model transfers to consistency with the final policy’s own previous plans? Evaluation with responsive surrounding vehicles would further test this claim.
4 comments

We thank the reviewer for the careful reading and for the constructive suggestions. All changes and additions are included in the revised PDF. Below we respond to selective consistency, the isolated Stage-2 loss and smoothing baselines, and transfer to the policy's own plans.

Q1 (frames needing little revision vs. a timely change)

Answer: Frames are split by whether the log trajectory itself departs from the previous plan (Tables 6 and 7), and safety is measured separately on frames whose predecessor is unsafe (Table 8, App. A.2).

Evidence: On Stable frames (90.08%), the transition reward raises EC from 70.98 to 75.98 at EP 93.24 versus 93.23. On Revised frames (9.92%), EC moves from 12.05 to 13.64, i.e., the plan still departs on 87.9% and 86.4% of these frames, while EP rises 96.10 to 96.28, NC stays 98.90, and TTC improves 98.49 to 98.59 (Table 6). Over the 154 frames whose predecessor is unsafe under both checkpoints, safe corrections rise 35 to 64, unsafe overturns fall 27 to 20, carried-over errors fall 53 to 38, the overturn rate rises 40.3% to 54.5%, safe-overturn accuracy 56.5% to 76.2%, and EP 94.27 to 94.57 (Table 8). On the EC of 69.8 in Table 2: methods with higher EC report lower ego progress (86.0 to 88.9 versus 93.4), and Revised frames have higher mean speed (6.08 versus 5.21 m/s) and mean EP (96.10 versus 93.24) than Stable frames (Table 7), so lower EC partly reflects more frequent revision at higher speed rather than worse execution. The reward thus suppresses changes only on stable frames, preserves them where the log indicates a revision, and improves replacement safety.

Boundary: Reaction delay is undefined under the non-reactive NAVSIM protocol, where agents replay logged trajectories and each frame starts from the logged ego state in an independent rollout; the collision outcome is the NC rate, which does not degrade (98.90 on Revised frames). NAVSIM also has no hazard annotations, so the log's own revision is the criterion for a required change. The proxies are whether the plan changes when the log indicates it should, and whether NC and TTC degrade; neither degrades.

Revision: Tables 6--8 (App. A.2) are added.

Q2 (isolating the Stage-2 loss; temporal smoothing under the same budget)

Answer: The loss is ablated with the diffusion expert fixed, and three temporal-consistency baselines are compared under the same architecture and budget.

Evidence: Table 3 splits Stage 2 into IDs 2a and 2b: 2a trains the diffusion expert alone (EC 62.15, EPDMS 90.10); 2b adds only the coherence term (LtcL_{tc}) with the expert unchanged (EC 65.10, EPDMS 90.35), so the loss contributes +2.95 EC on its own. Two baselines act at inference time on the per-frame checkpoint (temporal ensembling, hysteresis); the third retrains the same planner with the previous plan as an extra input. In Table 9, temporal ensembling reaches EPDMS 89.10 and EC 69.40, hysteresis 89.87 and 61.23, the history-conditioned baseline 90.35 and 65.10 with EP 92.57, the transition reward 90.94 and 69.80 with EP 93.40. Table 10 gives the sweep over α\alpha and δ\delta: at α=0.7\alpha = 0.7, temporal ensembling reaches EC 74.29, above ours, but EPDMS falls to 87.10, EP to 92.85, and Revised NC to 98.54. Temporal smoothing can therefore inflate EC alone by anchoring to the past; only the proposed objective raises EC without lowering EPDMS or progress.

Revision: Table 3 separates IDs 2a and 2b and includes EC; Tables 9, 10 (App. A.3) are added.

Q3 (transfer to the policy's own plans; responsive traffic)

Answer: We add two quantities to Table 5 and track one through training: ECrefEC_{\mathrm{ref}}, the indicator against a common frozen reference, and drefd_{\mathrm{ref}}, the normalized drift to it. Reactive-traffic validation is not possible on NAVSIM, which is non-reactive by construction, so no responsive-vehicle experiment is run and every method shares this protocol.

Evidence: ECrefEC_{\mathrm{ref}} orders variants identically to benchmark EC, computed between the policy's own consecutive plans: ECrefEC_{\mathrm{ref}} = 65.47, 64.52, 66.36 against EC = 69.80, 68.98, 72.12 for IDs 1--3, both giving ID 3 > ID 1 > ID 2. As ECrefEC_{\mathrm{ref}} uses one frozen reference for every row, the identical ordering shows that coherence against the frozen reference carries over to coherence between the policy's own plans. For the drift, drefd_{\mathrm{ref}} at the start of Stage 3 is 0.2522, 0.2676, 0.2752, varying by 9% across variants whose EC spans 68.98 to 72.12 and whose EP spans 92.27 to 93.51; tracking it per Stage-3 epoch gives an increase of at most 15% and a maximum of 0.32, so the gap stays bounded.

Boundary: The non-reactive protocol limits the claim to navtest consistency, not responsive traffic or closed-loop execution (Sec. 6).

Revision: Table 5 includes the drefd_{\mathrm{ref}} and ECrefEC_{\mathrm{ref}} columns for IDs 1--3; Sec. 5.3 reports the per-epoch drift.

Thank you for the additional experiments. I am still concerned about whether the consistency objective preserves timely responses to sudden hazards. The log-based analysis shows that the planner can revise its plans, but does not measure reaction delay during closed-loop execution. I am keeping my rating.

We sincerely thank the reviewer for the thoughtful follow-up and for taking the time to review our additional analyses.

Answer: To address the concern that the consistency objective may delay responses to sudden hazards, we add a closed-loop evaluation in HUGSIM, which captures dynamic responsiveness and real-time adaptation that the open-loop NAVSIM protocol cannot.

Evidence: HUGSIM is a closed-loop simulator built on reconstructed nuScenes, KITTI-360, Waymo and PandaSet scenes, designed for photo-realistic closed-loop evaluation with scripted and adversarial road users~\cite{zhou2025hugsim}; such actors make delayed hazard responses observable. The recorded surrounding traffic is replayed while the ego is driven by the evaluated planner, so the policy's decisions change the scene it subsequently observes. At the hard and extreme tiers, the injected road users are constant-velocity scripted vehicles and an ego-predicting adversarial planner, respectively. We evaluate the official HUGSIM benchmark of 436 scenarios (nuScenes 88, KITTI-360 113, Waymo 108, PandaSet 127), each at four difficulty tiers with identical recorded traffic. Both checkpoints are run on the same scenario files, initial states, injected-actor configurations, and replays, with one deterministic rollout per scenario, yielding 436 paired scenario-level comparisons per checkpoint; since the scenario files and injected-actor schedules are fixed before either checkpoint is run, hazards are not selected from the policy's behavior. We report HUGSIM's closed-loop metrics (NC, DAC, TTC, comfort, PDMS, RC, HD-Score) and per-scenario collision occurrence, comparing continuous metrics with paired Wilcoxon signed-rank tests and collision counts with McNemar's exact test. Reaction delay is measured from the scripted onset of the injected actor, taken from the scenario configuration, to the first step at which the planned waypoints depart from their pre-event trend, and is reported in planning steps and seconds.

Boundary: Because the recorded traffic is replayed, responsiveness to the ego comes from the injected actors; the reported latency is planner-level, and the runs are in progress, so the closed-loop values will be reported once they complete.

Revision: The HUGSIM experiment results will be reported in the revised PDF.

Reference: Hongyu Zhou, Longzhong Lin, Jiabao Wang, Yichong Lu, Dongfeng Bai, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. HugSim: A real-time, photo-realistic and closed-loop simulator for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.

We thank the reviewer for the follow-up and for acknowledging the additional analyses.

  • Answer: In closed-loop execution the transition reward improves riding comfort without a statistically detectable change in hazard response. On the 204 Medium+Hard pairs, the response rate is 65.4%65.4\% versus 66.2%66.2\% (McNemar p=0.79p = 0.79) and the median delay 2.88 s2.88\text{ s} versus 2.75 s2.75\text{ s} (Wilcoxon p=0.58p = 0.58; paired increase 0.12 s0.12\text{ s}, 95%95\% CI upper bound 0.50 s0.50\text{ s}), with NC 0.5170.517 versus 0.5210.521 (p=0.78p = 0.78) (Table 12) and comfort 0.9000.900 to 0.9380.938 (p<0.001p < 0.001) (Table 13). The ego-predicting tier, in which actors react to the ego, shows the same pattern: 49.4%49.4\% versus 50.5%50.5\% response rate and 2.25 s2.25\text{ s} delay in both cases (Table 12).
  • Evidence: HUGSIM runs closed-loop with the ego driven by the evaluated planner and extra road users injected as constant-velocity (Medium, Hard) or ego-predicting adversarial (Extreme) actors, with the per-frame and transition-reward checkpoints sharing scenario files, initial states and replays. Onset is the step at which an injected actor enters 20 m20\text{ m} of the logged ego reference path, and a response is a persistent two-step departure of the plan; per-tier values and the secondary metrics are in Tables 12–13 (App. A.5).
  • Summary: Read together, the closed-loop results show that the consistency objective delivers its gain as riding comfort, with the hazard response rate, delay and collision score unchanged across tiers.
  • Revision: Sec. 5.4 and App. A.5 with Tables 12–13 are added to the revised PDF.
Reviewer_HEqUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies consistency between consecutive plans in autonomous driving. It adds a differentiable consistency loss during imitation learning and a transition reward during reinforcement learning to a VLM-conditioned diffusion planner. The transition term compares the dynamics of consecutive trajectories over their overlapping horizon, using a frozen Stage 2 model to provide the previous-plan reference during reinforcement learning. The main contribution is this pairwise training objective and its asymmetric update scheme.

Critical strengths

The motivation makes sense to me. A planner can produce good individual trajectories while still changing its decisions unnecessarily across replanning steps. Treating this relation as a training objective addresses a meaningful problem, and the idea could be useful beyond the particular model used in this paper. 2. Table 4 provides encouraging evidence for the transition reward. EC improves from 65.13 to 69.80 and EPDMS improves from 90.45 to 90.94, while ego progress stays almost unchanged. This is a useful controlled comparison showing that the reward can improve the measured consistency without an obvious loss in average progress.

Critical weaknesses
  1. My main concern is that the framework is evaluated within a rather complicated system with strong pretrained components, which makes its own contribution less clear. Table 4 does isolate the benefit of the RL transition reward, but Table 3 introduces the diffusion expert and the imitation consistency loss together. The effect of the full proposed objective is therefore not well separated from the other changes. A controlled study with a simpler planner would make the contribution more convincing and help show whether the framework is broadly useful.
  2. The claims about preserving necessary changes and assigning inconsistency to the responsible decision need more support. EC uses fixed thresholds, and the frozen predecessor encourages agreement with an older policy. These are reasonable regularization choices, but they do not directly determine whether a change is justified or which decision was wrong. Table 2 also reports an EC of 69.8, below most compared methods, despite the highest overall score. This does not invalidate the approach, since necessary changes may reduce EC, but the current results do not show whether this explains the observed tradeoff.
Questions for authors

. For me, the main question is how much the proposed framework helps without such a heavy model. Table 4 is helpful for understanding the RL reward, but could the authors also separate the imitation consistency loss from the diffusion expert? Even a small experiment on a simpler planner would make the contribution easier to see. 2. I was a bit confused about where the previous plan comes from. During RL it is provided by the frozen Stage 2 model, but what is used when Stage 2 itself is being trained? And at evaluation, are both plans generated by the final policy? If so, I would like to understand how this difference affects the method. Related to this, what exactly changes when stop-gradient is removed in Table 5? Since EC is a non-differentiable reward, the update to the previous decision needs some explanation. 3. It would also help to see results for situations where the plan should change, such as a sudden hazard or an incorrect previous plan, alongside situations where the scene stays stable. The current averages leave me unsure whether the method handles these cases differently, especially when ego progress is not matched.

2 comments

We thank the reviewer for the constructive assessment. All changes and additions are included in the revised PDF. Below we separate the contribution, clarify the reference and stop-gradient, and show behavior when a change is needed.

W1/Q1 (separating the contributions; simpler planner)

Answer: Two results isolate the objective: the coherence term is ablated separately from the diffusion expert, and the objective is tested on a fully trained planner without a VLM.

Evidence: Table 3 splits Stage 2 into IDs 2a and 2b: 2a trains the diffusion action expert alone (EC 62.15, PDMS 94.06, EPDMS 90.10); 2b adds only the coherence term (LtcL_{tc}) with the expert unchanged (EC 65.10, PDMS 94.17, EPDMS 90.35), contributing +2.95 EC on its own. The objective is also applied to DiffusionDrive at its released, fully trained checkpoint (Table 11, App. A.4). RL raises PDMS to 91.2, EC 87.7 to 91.0, EPDMS 84.5 to 85.5; the transition reward further raises PDMS to 91.8, EC to 92.6, EPDMS to 85.8, EP 88.9 to 89.1, with no metric degrading. Both RL rows share the frozen checkpoint, samples, and compute budget, differing only in the transition term, so 91.2 to 91.8 is attributable to the objective, not to extra training; the 88.1 to 91.2 rise reflects RL extracting further performance from a converged planner, which imitation on that checkpoint cannot provide. Neither gain is confounded with the starting point: Table 11 fixes the base, and Table 4 gives the same control at the stronger base. The gain is smaller because EC already starts at 87.7, but the direction is consistent, so the objective is not tied to the VLA architecture.

Revision: Table 3 now separates IDs 2a (expert only) and 2b (expert plus LtcL_{tc}) and includes EC; Table 11 and App. A.4 are added.

Q2 (reference, evaluation, stop-gradient)

Answer: The stages use different references: during imitation, the logged expert plan (the only ground truth); during reinforcement, the frozen Stage-2 plan (a fixed, simulator-verifiable anchor with no gradient path to the policy); at evaluation, the final checkpoint, so consistency is measured between the policy's own predictions. In all three cases only the current decision is updated; the choice changes what is measured, not who is corrected. Removing the stop-gradient changes this: the predecessor becomes trainable, the update also lowers the likelihood of its own forward pass and couples successive decisions, raising EC but degrading EP and EPDMS.

Evidence: During Stage 2 the reference is the expert plan of the preceding log frame, cached offline in the current ego frame: a fixed tensor, not a prediction; during reinforcement, the frozen Stage-2 plan. At evaluation both plans come from the final checkpoint in two independent passes; the policy carries no cross-frame state, so the training reference does not leak into inference. EC is returned by the simulator after rolling out both plans, so it carries no gradient and the stop-gradient is a no-op for the cached and frozen references. With the policy's own previous plan, removing it (Table 5, ID 3) adds a gradient path to that forward pass: EC rises to 72.12, but EP falls to 92.27 and EPDMS to 90.71, against EC 68.98, EP 93.51, and EPDMS 90.78 for ID 2, which keeps the stop-gradient with the same reference.

Clarification: Inconsistency is charged only to the decision that produced it; fixing the reference does this in all cases, and the stop-gradient enforces it when the reference is self-referential.

Revision: Sec. 4.2 states the reference choice and Sec. 4.3 the stop-gradient rationale; Table 5, IDs 2--3 give the measurements.

Q3 (frames that should change; progress-matched comparison; EC 69.8)

Answer: The evaluation is split by whether the log revises the plan and whether the predecessor is unsafe, and the comparisons are progress-matched.

Evidence: Table 6 splits navtest by frame type. On Stable frames (90.08%, Table 7), EC rises 70.98 to 75.98 at essentially unchanged EP (93.24 versus 93.23). On Revised frames (9.92%), EC moves only 12.05 to 13.64, i.e., the plan still departs on 87.9% and 86.4% of them, while EP rises 96.10 to 96.28, NC stays 98.90, and TTC improves 98.49 to 98.59. Table 8 (App. A.2) over the 154 frames unsafe under both checkpoints: safe corrections 35 to 64, unsafe overturns fall 27 to 20, carried-over errors fall 53 to 38, overturn rate 40.3% to 54.5%, safe-overturn accuracy 56.5% to 76.2%, EP 94.27 to 94.57. On the EC of 69.8 in Table 2: methods with higher EC report lower ego progress (86.0 to 88.9 versus 93.4), and Revised frames have higher mean speed (6.08 versus 5.21 m/s) and mean EP (96.10 versus 93.24) than Stable frames, so a lower EC partly reflects more frequent revision at higher speed, not worse execution.

Boundary: NAVSIM has no hazard annotations; a required change is defined by the log's revision and the safety check on the predecessor.

Revision: Tables 6, 7, and 8 (App. A.2) are added.

We would also like to clarify to the reviewer that, in response to the concern that the consistency objective may delay responses to sudden hazards, we have added a closed-loop hazard evaluation to the revised PDF. This aims to measure, through closed-loop simulation rather than log-based proxies, whether the transition reward preserves timely hazard responses while improving riding comfort.

  • Answer: The closed-loop evaluation shows that the transition reward improves riding comfort without a statistically detectable change in hazard response. On the 204 Medium+Hard pairs, the response rate is 65.4%65.4\% versus 66.2%66.2\% (McNemar p=0.79p = 0.79) and the median delay 2.88 s2.88\text{ s} versus 2.75 s2.75\text{ s} (Wilcoxon p=0.58p = 0.58; paired increase 0.12 s0.12\text{ s}, 95%95\% CI upper bound 0.50 s0.50\text{ s}), with NC 0.5170.517 versus 0.5210.521 (p=0.78p = 0.78) (Table 12), and comfort rises from 0.9000.900 to 0.9380.938 (p<0.001p < 0.001) (Table 13).
  • Evidence: The per-frame and transition-reward checkpoints are both driven closed-loop in HUGSIM with injected constant-velocity (Medium, Hard) and ego-predicting adversarial (Extreme) actors, sharing scenario files, initial states and replays; per-tier values and the secondary metrics are in Tables 12–13 (Sec. 5.4, App. A.5).
  • Summary: Taken together, the coherence gain is realized as riding comfort, while the hazard response rate, delay and collision score remain unchanged.
  • Revision: Sec. 5.4, App. A.5, and Tables 12–13 are added to the revised PDF.
Reviewer_TGHKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ReTranDrive aims to make autonomous driving plans more consistent across planning steps. It compares the overlapping parts of consecutive plans using acceleration, jerk, yaw rate, and yaw acceleration. The method adds a consistency loss during imitation learning and a transition reward during RL. During RL, a frozen Stage-2 plan serves as the reference, so the consistency-related update acts only on the current plan. Experiments on NAVSIM show strong planning results, with ablations supporting the reward and reference designs.

Critical strengths
  1. The paper studies an important problem: a driving model may produce good individual plans but keep changing its decisions across planning steps.

  2. The method includes consistency objectives in both imitation learning and RL. Comparing the overlapping parts of consecutive plans gives a clear way to measure these changes.

  3. The NAVSIM results are strong, and the ablations test different reward forms and reference choices. In Table 4, the transition reward improves EC from 65.13 to 69.80, with almost no change in ego progress.

Critical weaknesses
  1. Earlier work, such as MomAD, already uses past plans to improve temporal stability. ReTranDrive adds different training objectives, but its conceptual novelty appears incremental. A comparison with a simple history-conditioned baseline would help show the benefit of the proposed objective.

  2. The consistency measure uses fixed thresholds and does not directly check whether a plan change is needed because the scene has changed. The authors explain that other reward terms can support necessary changes, but there are no targeted tests showing whether the method reacts quickly enough to sudden hazards.

  3. Table 5 supports the use of a frozen Stage-2 reference. However, this reference differs from the previous plan produced by the updated Stage-3 model. The paper does not show how this gap changes during training or how it affects consistency between the model’s own consecutive plans.

  4. Table 4 provides quantitative evidence of better consistency, but Figures 2–4 mainly show individual trajectories. They do not directly show the overlapping parts of consecutive plans or how their acceleration, jerk, and turning behavior change.

Questions for authors
  1. How does ReTranDrive compare with a baseline that uses the same model but takes the previous plan as an extra input, under the same training setup?

  2. Could the authors test sudden cut-ins, emergency braking, and unexpected obstacles, with and without the transition objective? Please compare reaction time and collision rate.

  3. Beyond the comparison in Table 5, how does the gap between the frozen Stage-2 reference and the Stage-3 policy change during training? Does consistency with the frozen reference also lead to consistency between consecutive Stage-3 plans?

  4. Please show time-aligned pairs of consecutive plans and plot their differences in acceleration, jerk, and turning behavior. Examples where a sudden hazard requires a large plan change would be especially helpful.

2 comments

We thank the reviewer for the constructive comments. All changes and additions are included in the revised PDF. Answers point to material added in the discussion (Tables 6--10, Fig. 5).

W1/Q1 (novelty; history-conditioned baseline)

Answer: MomAD (Song et al., CVPR 2025) is not evaluated on NAVSIM, so its scores are not comparable. We instantiate the same idea (using the previous plan to stabilize behavior) under our setting; it does not reproduce the gain of our objective.

Evidence: Table 9 reports the retrained history-conditioned baseline (previous plan as an extra input, same budget): EPDMS 90.35, EC 65.10, EP 92.57, against 90.94, 69.80, 93.40 for the transition reward; Stable-frame EC 70.88 against 75.98. Inference-time variants also fall short: temporal ensembling reaches EPDMS 89.10 and EC 69.40, hysteresis 89.87 and 61.23; in Table 10, α=0.7\alpha = 0.7 gives EC 74.29 but lowers EPDMS to 87.10, EP to 92.85, and Revised NC to 98.54, so anchoring to the past inflates EC alone. The difference is where the past plan enters: baselines add it as an input or anchor while the training signal stays per-frame; ReTranDrive places it in the objective and charges inconsistency to the decision that produced it. Table 3 separates the components: the coherence term raises EC from 62.15 (ID 2a) to 65.10 (ID 2b), and the transition reward reaches 69.80 (ID 3).

Revision: Sec. 2.3 now contrasts conditioning on the historical plan with optimizing the transition; the baseline is added in Table 9 (App. A.3).

W2/Q2 (necessary changes; hazards)

Answer: Two tests show the response is preserved and improves: changes are not suppressed when required, and unsafe predecessors are more often replaced by safe plans.

Evidence: Table 6 groups frames by whether the log revises the plan. On Revised frames (9.92%, Table 7), EC moves only from 12.05 to 13.64: the plan departs on 87.9% and 86.4% of them at the same step at which the log changes, so the response occurs within one step; NC stays 98.90, TTC improves 98.49 to 98.59, EP rises 96.10 to 96.28. On Stable frames EC rises from 70.98 to 75.98 with EP unchanged. Table 8 (App. A.2), over the 154 frames unsafe under both checkpoints: safe corrections 35 to 64, unsafe overturns 27 to 20, carried-over errors 53 to 38, overturn rate 40.3% to 54.5%, safe-overturn accuracy 56.5% to 76.2%, EP 94.27 to 94.57. Rows differ only in the transition term, giving the with/without comparison.

Boundary: Labeled cut-in and braking tests are impossible in NAVSIM: it is non-reactive, agents replay logged trajectories, and no hazard annotations exist, so closed-loop reaction delay is undefined; the same-step departure and NC/TTC are the proxies (Sec. 6).

Revision: Tables 6--8 (App. A.2) are added.

W3/Q3 (transfer to own previous plans)

Answer: Consistency against the frozen reference transfers to the policy's own consecutive plans, and the drift grows slowly and stays bounded.

Evidence: Table 5 reports, for IDs 1--3, the drift dˉref\bar d_{\mathrm{ref}} and ECrefEC_{\mathrm{ref}} at the start of Stage 3: 0.2522, 0.2676, 0.2752 and 65.47, 64.52, 66.36, against benchmark EC 69.80, 68.98, 72.12. As ECrefEC_{\mathrm{ref}} uses one common reference, it orders the variants identically to EC. Mean dˉref\bar d_{\mathrm{ref}} over Stage-3 epochs rises by at most 15% (max 0.3162).

Boundary: Evaluation follows the non-reactive NAVSIM protocol, so responsive-traffic benefits are not measured; all methods share this setting (Sec. 6).

Revision: Table 5 gains the two columns; Sec. 5.3 reports the per-epoch drift.

W4/Q4 (time-aligned plan pairs)

Answer: Fig. 5 (App. A.1) shows time-aligned consecutive plans and their per-step kinematic differences, for a frame that should stay consistent and one that needs a change.

Evidence: The upper row plots Yk−1ovY_{k-1}^{\mathrm{ov}} and YkovY_k^{\mathrm{ov}} over the shared horizon; the lower rows plot per-step differences of acceleration, jerk, yaw rate, and yaw acceleration normalized by θq\theta_q, with the ±θq\pm\theta_q band marking admissible deviation. On the Stable frame the plans nearly coincide (max gap 0.62 m) and the ratios (0.17, 0.31, 0.45, 0.39) stay inside the band. On the Revised frame the max gap is 13.57 m, the ratios rise to 1.49, 1.13, 1.59, 1.80, all four channels leave the band together, and the model departs in the same direction as the human plan (lateral offset 0.9 to 4.84 m).

Boundary: The large-change example is the frame the log itself revises into a lane change; NAVSIM has no hazard annotations, so a labeled cut-in cannot be drawn and the log's revision defines when a change is required.

Revision: Fig. 5 and the text in App. A.1 are added.

In summary, the history-conditioned baseline and two heuristics fall below the transition objective (Table 9); Revised frames and unsafe predecessors show no degradation in NC, TTC, or EP (Tables 6, 8); and ECrefEC_{\mathrm{ref}} orders variants as EC does while dˉref\bar d_{\mathrm{ref}} stays bounded (Table 5).

We would also like to clarify to the reviewer that, in response to the concern that the consistency objective may delay responses to sudden hazards, we have added a closed-loop hazard evaluation to the revised PDF. This aims to measure, through closed-loop simulation rather than log-based proxies, whether the transition reward preserves timely hazard responses while improving riding comfort.

  • Answer: The closed-loop evaluation shows that the transition reward improves riding comfort without a statistically detectable change in hazard response. On the 204 Medium+Hard pairs, the response rate is 65.4%65.4\% versus 66.2%66.2\% (McNemar p=0.79p = 0.79) and the median delay 2.88 s2.88\text{ s} versus 2.75 s2.75\text{ s} (Wilcoxon p=0.58p = 0.58; paired increase 0.12 s0.12\text{ s}, 95%95\% CI upper bound 0.50 s0.50\text{ s}), with NC 0.5170.517 versus 0.5210.521 (p=0.78p = 0.78) (Table 12), and comfort rises from 0.9000.900 to 0.9380.938 (p<0.001p < 0.001) (Table 13).
  • Evidence: The per-frame and transition-reward checkpoints are both driven closed-loop in HUGSIM with injected constant-velocity (Medium, Hard) and ego-predicting adversarial (Extreme) actors, sharing scenario files, initial states and replays; per-tier values and the secondary metrics are in Tables 12–13 (Sec. 5.4, App. A.5).
  • Summary: Taken together, the coherence gain is realized as riding comfort, while the hazard response rate, delay and collision score remain unchanged.
  • Revision: Sec. 5.4, App. A.5, and Tables 12–13 are added to the revised PDF.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.