acceptodds
Under review as a conference paper at ICLR 2027

DriveAnchor: Progressive Anchor-based Flow Learning for Autonomous Driving Planning

12 views

Abstract

Autonomous-driving planners must generate route- and maneuver-compliant trajectories under a limited decoder-call budget while adapting offline to black-box static-collision feedback. We propose DriveAnchor, a three-stage framework built around a shared vocabulary of recorded trajectories and a continuous trajectory interface. The vocabulary supports flow-matching (FM) residual learning, Energy Field (EF) corridor-conditioned initialization, and anchor-guided zeroth-order reward fine-tuning. The reward update operates in trajectory space without policy likelihoods or reward gradients, enabling offline adaptation of a deterministic decoder. On sampled subsets of nuPlan Val14 and Test14-hard per simulation mode, DriveAnchor achieves the highest reactive scores among the compared methods: 81.50 and 65.74, respectively. We further integrate the trajectory generator into a real vehicle's planning pipeline and demonstrate its operation on urban roads. On 10,000 held-out internal scenarios, under the original non-EF prior and a fixed two-call FM budget, reward fine-tuning reduces short- and long-horizon static candidate collision rates from 27.22% and 56.54% to 2.94% and 7.15%, respectively, and improves 8-s best-of-set minADE from 2.27 to 1.34 m. In a separate junction diagnostic on checkpoints before Stage 3, one and two EF calls raise corridor-guidance satisfaction from 22.10% to 74.13% and 87.92%, respectively. Additional ablations support the benefits of real-trajectory priors and anchor-guided exploration. Code and deployment videos are available at the anonymous project page: https://driveanchor.github.io/. In real-vehicle tests, trajectory generation takes 2.06 ms on average.

Primary area:
applications to robotics, autonomy, planning
Keywords:
autonomous driving, controllable trajectory generation, flow matching, zeroth-order optimization, reward fine-tuning

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_EadHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DriveAnchor, a three-stage autonomous driving planning framework built on a real-trajectory vocabulary: first, it learns trajectory residual correction via flow matching; then it uses road-corridor conditioning to guide trajectory initialization; finally, it applies trajectory-anchor-guided zeroth-order optimization to incorporate non-differentiable reward feedback for fine-tuning the deterministic generator. The authors validate the method on nuPlan, an internal dataset, and real-vehicle scenarios. Overall, the main contribution of this work lies in the integration of the planning system and engineering practice, while the innovation at the machine learning methodological level is relatively limited.

Critical strengths
  1. Real-vehicle deployment demonstrates good engineering practicality. The authors show scenarios such as lane changes, pedestrian interactions, intersection turns, and construction zones, and report trajectory generation efficiency on an onboard Orin platform. These results demonstrate the ability to integrate generative planning methods into an actual driving system, which is the most commendable part of this paper.

  2. Black-box reward adaptation has practical value, and the internal experiments also show clear gains. For deterministic generators without an explicit probability density, using external rewards to improve candidate trajectories has a reasonable application motivation. Table 2 shows that reward fine-tuning can significantly reduce the static collision ratio among candidate trajectories; corridor guidance also improves the corresponding geometric compliance metrics. These results support the practical value of the framework to some extent.

Critical weaknesses
  1. Limited ML novelty; necessity of the complex design is insufficiently justified. The three stages are largely a task-specific combination of existing ideas (trajectory priors, conditional guidance, reward optimization). The trajectory-vocabulary search direction in Stage 3 is somewhat distinctive, but the paper does not clearly articulate what generalizable learning mechanism or new insight it provides, nor does it show that the full pipeline is clearly superior to simpler alternatives with comparable training budgets.

  2. Insufficient benchmark coverage and strong-baseline comparison. nuPlan experiments cover only sampled Val14 and Test14-hard scenarios and omit Test14. Diffusion Planner and Flow Planner are cited but not included in the main closed-loop comparison; public reports show their non-reactive Test14-hard scores without refinement are 75.99 and 76.47, versus 63.37 here. The authors need to clarify the core contribution given this clear SOTA gap. Diffusion Planner, Flow Planner

  3. Candidate-level gains are not shown to translate into final closed-loop driving benefits. Ablations mainly focus on candidate-set static collision ratio, minimum trajectory error, and corridor compliance. But actual driving executes one selected trajectory; a better candidate set does not necessarily imply safer or more efficient final driving. The paper lacks a unified closed-loop ablation of the three stages, making it hard to assess each stage’s independent contribution and whether the gains justify the added complexity.

  4. Lack of systematic closed-loop validation in high-interaction scenarios. The paper emphasizes closed-loop planning but does not evaluate on interaction-rich benchmarks such as Bench2Drive, which explicitly tests overtaking, merging, yielding, and emergency braking. Since the main ablations concern static collisions and corridor compliance, it remains unclear whether these improvements transfer to complex interactive decision-making. The authors should add such evaluation or justify why the method is unsuitable for Bench2Drive. Bench2Drive

Overall, I appreciate the value of this paper in autonomous driving system integration and real-vehicle deployment, and its contribution form is closer to application and system research in robotics or intelligent transportation. Measured by ICLR's requirements for machine learning contributions, the current methodological novelty and mainstream benchmark evidence are still insufficient to form a convincing case for acceptance. IMHO, in terms of either algorithmic innovation or benchmark metrics, at least one of them should be sufficiently outstanding.

Questions for authors
  1. Limited ML novelty; necessity of the complex design is insufficiently justified. The three stages are largely a task-specific combination of existing ideas (trajectory priors, conditional guidance, reward optimization). The trajectory-vocabulary search direction in Stage 3 is somewhat distinctive, but the paper does not clearly articulate what generalizable learning mechanism or new insight it provides, nor does it show that the full pipeline is clearly superior to simpler alternatives with comparable training budgets.

  2. Insufficient benchmark coverage and strong-baseline comparison. nuPlan experiments cover only sampled Val14 and Test14-hard scenarios and omit Test14. Diffusion Planner and Flow Planner are cited but not included in the main closed-loop comparison; public reports show their non-reactive Test14-hard scores without refinement are 75.99 and 76.47, versus 63.37 here. The authors need to clarify the core contribution given this clear SOTA gap. Diffusion Planner, Flow Planner

  3. Candidate-level gains are not shown to translate into final closed-loop driving benefits. Ablations mainly focus on candidate-set static collision ratio, minimum trajectory error, and corridor compliance. But actual driving executes one selected trajectory; a better candidate set does not necessarily imply safer or more efficient final driving. The paper lacks a unified closed-loop ablation of the three stages, making it hard to assess each stage’s independent contribution and whether the gains justify the added complexity.

  4. Lack of systematic closed-loop validation in high-interaction scenarios. The paper emphasizes closed-loop planning but does not evaluate on interaction-rich benchmarks such as Bench2Drive, which explicitly tests overtaking, merging, yielding, and emergency braking. Since the main ablations concern static collisions and corridor compliance, it remains unclear whether these improvements transfer to complex interactive decision-making. The authors should add such evaluation or justify why the method is unsuitable for Bench2Drive. Bench2Drive

11 comments
AuthorsOpenReview ↗

Response to Reviewer EadH (Part 1/2): Contribution and Benchmark Positioning (W1/Q1, W2/Q2)

We thank the reviewer for recognizing the practical value of the real-vehicle deployment and of black-box reward adaptation.

W1 & Q1: Limited machine-learning novelty; the necessity of the three-stage design.

A1. Our contribution is a compositional planning framework built for real-vehicle deployment, with evidence for each component. The three components work on a shared trajectory representation and solve different problems:

(i) FM generates the candidates. A vocabulary of real trajectories supplies initial trajectories, and FM learns scene-conditioned continuous residuals.

(ii) EF applies geometric requirements such as corridors. Corridor compliance of the final output rises from 62.02% to 73.42% (Sec. 4.5). In the new nuPlan closed-loop runs (tables in Part 2), the overall score is 74.08 with EF and 72.28 without it; removing EF lowers the making-progress, route-progress, and comfort sub-scores by 1.02, 0.88, and 1.02 points.

(iii) RL improves candidates from black-box feedback. With FM×2 fixed, candidate static collision rates fall from 27.22% / 56.54% to 2.94% / 7.15% and minADE@80 from 2.27 m to 1.34 m (Table 2 of the paper). In the new nuPlan runs, two epochs of RL raise Test14-hard NR / R by 2.89 / 0.94 points and the overall score from 74.08 to 74.55; the no-at-fault-collision and TTC sub-scores rise by 1.40 and 1.27 points. On 21,082 internal cases, the pass rate rises from 44.78% to 45.80%.

(iv) Real-vehicle deployment. To balance trajectory diversity and the real-time compute budget, the on-vehicle system randomly samples six anchors from the vocabulary, generates six candidates, and uses a rule-based selector to choose the executed trajectory. The system operates on thousands of autonomous vehicles in urban areas across dozens of cities, with millions of kilometers of autonomous driving accumulated. The autonomous-driving share is 99%.

Stage 3 supports efficient and extensible reward adaptation. After generating a center trajectory, we construct exploration probes directly in output space and query their rewards without rerunning the generator for each probe. The vocabulary, neighbor-selection rule, and perturbation scale provide explicit interfaces for incorporating motion constraints and driving preferences while retaining the generator architecture and decoding procedure. With the same initialization and query budget, standard ZO gives 1.97% all-candidate static collision vs. 0.28% for ours (Appendix I.3, Table 8).

The adaptation stages reuse the existing generator and frozen scene encoder. Stages 1, 2, and 3 use 120M, 7.5M, and 2.5M samples and cost 902, 541, and 243 GPU·hours (Table 5). On nuPlan, the base FM uses about 21k samples; EF adaptation and the scoring head use 4,608 of them. This organization supports targeted updates to an existing planner with an explicit additional training budget.

The contribution combines a shared trajectory interface, measurable component effects, and execution-level evidence from public and internal closed-loop tests and real deployment. This deployment-oriented scope is consistent with the primary area under which the paper was submitted: "applications to robotics, autonomy, planning".

W2 & Q2: Limited benchmark coverage; a gap to the state of the art.

A2. We study controllable generation and feedback adaptation for real-world operation, and nuPlan serves to validate the approach under a public protocol.

On Test14-hard NR, Diffusion Planner and Flow Planner report non-refined scores of 75.99 and 76.47. Our paper reports 63.37 on its sampled subset (Table 1). On the common valid scenarios used in this round, BASE scores 64.01 and the model after two RL epochs scores 66.90.

There is also a sim-to-real gap: perception errors, the behavior of other traffic participants, control tracking errors, and on-board real-time constraints are not fully reproduced by any single simulation protocol. We therefore read public evaluation, internal closed-loop testing, and real-vehicle operation together.

Our public study examines the additional cost and benefit of adapting an existing generator. The base FM uses about 21k nuPlan samples for 3 epochs; EF adaptation and the scoring head use 4,608 of them for 7 and 12 epochs; this round adds 2 RL epochs. Diffusion Planner and Flow Planner report a different training regime: about one million scenarios, with 500 and more than 200 epochs, respectively. For a deployed planner, a practical question is how an existing model can be updated through a short training cycle targeting newly observed failures. The public EF/RL comparisons evaluate this incremental adaptation capability alongside base planning performance. The original model is on par with PlanTF on Val14 NR (84.80 vs. 84.72) and has the highest R score among the compared methods on both sampled subsets.

AuthorsOpenReview ↗

Response to Reviewer EadH (Part 2/2): Closed-Loop Ablations and Interactive Scenarios (W3/Q3, W4/Q4)

W3 & Q3: Do candidate-level improvements translate into closed-loop gains? Component ablations are missing.

A3. We add public closed-loop comparisons for EF and RL.

The original public BASE model uses Stage 1 + Stage 2 without Stage 3; the new RL model adds two epochs of reward adaptation. All three configurations keep the scoring head and execution post-processing and use 3,000 candidates and FM×2 (EF×2 when enabled), on the paper's scenario lists.

ConfigurationVal14 NRVal14 RTest14-hard NRTest14-hard ROverall
BASE: EF + scoring head84.8981.3164.0165.9574.08
No EF81.3479.1364.1364.3572.28
RL83.1181.1566.9066.8974.55

Common valid counts: 99/99/99/97 (394 total). Overall is weighted by these counts. The same exception tokens are excluded across configurations; completed zero scores are kept. Deltas use unrounded means.

EF adds 1.80 points to the pooled score. Two RL epochs change Test14-hard NR/R by +2.89/+0.94 points and Val14 NR/R by −1.79/−0.15 points, yielding +0.47 points overall. Sub-metric means on the reported scenarios (0–100, higher is better):

Sub-metricBASENo EFRLNo EF − BASERL − BASE
No at-fault collision92.7792.8994.16+0.13+1.40
Drivable-area compliance96.4596.1996.95−0.25+0.51
Making progress91.8890.8691.62−1.02−0.25
Progress along route67.0166.1467.39−0.88+0.37
Driving direction98.7398.6098.98−0.13+0.25
Time to collision (TTC)87.3187.3188.580.00+1.27
Speed limit99.2899.2399.28−0.050.00
Comfort91.1290.1091.37−1.02+0.25

The RL gains are concentrated in the safety-related items (collision +1.40, TTC +1.27, drivable area +0.51) while mean route progress changes by +0.37 and making progress by −0.25; on Test14-hard NR these three safety items are +2.53, +3.03, and +1.01. This matches the Stage 3 reward, which penalizes collisions with static obstacles and road boundaries. Removing EF mainly lowers making progress, route progress, and comfort. These comparisons evaluate the closed-loop task after downstream selection and execution, which directly addresses your question.

In addition, the internal closed-loop evaluation before and after RL covers 21,082 cases: passing cases rise from 9,441 to 9,656 and the pass rate from 44.78% to 45.80%; the pass rates for nudging and lane changing rise by 3.00 and 2.03 percentage points.

W4 & Q4: No closed-loop validation in complex interactive scenarios; Bench2Drive is suggested.

A4. The existing evidence already covers interactive closed-loop scenarios; Bench2Drive would be additional validation.

nuPlan includes non-reactive and reactive (R) settings; Val14 covers 14 scenario types, and Test14-hard consists of hard scenarios selected from them. The internal evaluation covers nudging, lane changing, merging, and intersection turns, and also overtaking, yielding, and emergency braking. Real-vehicle operation across cities further supports usability in a production system.

Bench2Drive offers a complementary CARLA-based evaluation of sensor-to-control driving. Integrating our planning module would add a perception front end and adapt the map/navigation and control interfaces, evaluating the combined system. The current Stage 3 reward targets static obstacles and road boundaries. In our deployment workflow, reward adaptation supports short updates targeting specific failure types and can be extended through other black-box rewards. Real-vehicle videos illustrating interactive situations can also supplement the anonymous project website.

The public reactive evaluation and internal interactive cases address executed behavior in the evaluated planning setting, while real-vehicle deployment demonstrates practical integration. Bench2Drive would extend this evidence to another sensor-to-control pipeline.

If these responses and the additional experiments have changed your view, we hope you will consider raising your score. Thank you for your time and your review.

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer,

Thank you again for your detailed and constructive review. In our response, we have clarified the positioning of this work as controllable trajectory generation for real-vehicle deployment and short-cycle black-box reward adaptation of an existing planner. We have also explained the distinct role that each of the three stages plays on the shared trajectory representation, and discussed how our training scale and setting differ from those of Diffusion Planner and Flow Planner.

In addition, we have added closed-loop component ablations of EF and RL on nuPlan, with the candidate count, decoder-call budget, and scoring head held fixed, to examine whether candidate-level improvements translate into execution-level benefits. We have further included a large-scale internal closed-loop evaluation covering interactive scenarios such as nudging, lane changing, merging, intersection turns, overtaking, yielding, and emergency braking, and clarified the relationship between Bench2Drive and our setting.

As the discussion period has now passed its halfway point, we would be grateful to know whether these additional results and clarifications have adequately addressed your main concerns. If any issue remains unresolved, or if any additional experimental details or analyses would be helpful for your assessment, please feel free to let us know. We would be very happy to provide further clarification and continue the discussion.

Thank you again for your time and consideration.

Best regards,

The Authors

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

DriveAnchor integrates anchor-based residual trajectory learning, corridor-conditioned Energy Field initialization, and offline zeroth-order reward fine-tuning. A shared trajectory vocabulary supports generation, control, and exploration without requiring differentiation of a static-collision reward. The paper reports reactive nuPlan improvements on sampled scenario subsets, substantial reductions in internal candidate collision rates, and real-vehicle integration. Its analysis explicitly distinguishes the local smoothed surrogate optimized by the update from the original reward objective.

Critical strengths

(+) The shared continuous trajectory interface gives the three stages a clear connection. Anchor-guided perturbations are a sensible way to obtain structured exploration directions even when a deterministic decoder produces limited output diversity.

(+) The internal results demonstrate a substantial effect at a fixed decoding budget. In Table 2, two-call reward fine-tuning reduces long-horizon static candidate collision from 56.54% to 7.15% and improves 8-second minADE from 2.27 to 1.34 m.

(+) The paper is explicit about several important qualifications: repeated calls reset the time input, the estimator targets a query-local surrogate, the reward is static, and the 2.06 ms Orin measurement excludes the encoder and selector. Real-vehicle integration adds practical interest.

Critical weaknesses

(-) The public evaluation leaves uncertainty about the stability and generality of the reported gains. Appendix A evaluates 100 sampled scenarios per benchmark and simulation mode, with the same fixed scenario list shared across methods. This supports comparisons on the selected subsets, but the rankings may depend on scenario selection. DriveAnchor achieves the highest reactive scores among the evaluated configurations, whereas its Test14-hard non-reactive score is lower than those of PlanTF and the reported PLUTO configuration. Larger evaluation sets or repeated sampling, together with paired confidence intervals, would help establish the robustness of these findings.

(-) The contribution of the individual components to public closed-loop performance remains insufficiently isolated. The internal ablations demonstrate substantial reductions in candidate-level static collision rates, but these metrics do not directly establish improvements in the single executed trajectory or its interactions with moving agents. Moreover, the public nuPlan configuration uses a score head for trajectory selection, whereas the internal studies evaluate candidate quality before downstream selection, including after residual screening. Closed-loop ablations of EF initialization and Stage 3 fine-tuning on nuPlan, with the candidate count, decoder-call budget, and selector held fixed, would more directly connect the internal improvements to the reported planning gains.

Questions for authors
  1. What are the public nuPlan scores before and after Stage 3 with the same candidate count, selector, and inference budget?

  2. Can the authors report complete pipeline latency and quantitative real-vehicle exposure, intervention, and failure statistics?

2 comments
AuthorsOpenReview ↗

Response to Reviewer UaLS (Part 1/2): Evaluation Scale and Closed-Loop Component Ablations (W1, W2/Q1)

We thank the reviewer for recognizing the shared trajectory interface, anchor-guided exploration, and the improvements under a fixed decoding budget. We respond below to the points on the stability of the public results, the link between components and execution-level benefit, and the deployment statistics.

W1: The public evaluation is small in scale; the stability and generality of the results need further support.

A1. The public comparisons use fixed sampled scenario lists, complemented by larger-scale internal closed-loop evaluation.

Each public setting uses a fixed list of 100 sampled scenarios. The comparisons concern these lists; the common valid cases for the additional configurations are specified below.

The lists are sampled with a fixed random seed (Appendix A), independently of any method's scores. Methods and additional configurations share these lists, enabling paired scenario-level comparisons. Complementary evidence comes from 21,082 internal closed-loop cases and candidate analysis on 10,000 scenes, derived from real-vehicle data and covering diverse driving situations.

The reviewer notes Test14-hard NR scores of 69.70 for PlanTF and 70.03 for PLUTO, compared with 63.37 in our paper. On the common valid scenarios used in this round, BASE scores 64.01 and the model after two RL epochs scores 66.90, a gain of 2.89 points from a short adaptation stage. Our public study examines both base planning performance and the additional execution-level benefit of feedback adaptation. The original Val14 NR scores are 84.80 for our model and 84.72 for PlanTF.

W2 & Q1: Do candidate-level improvements carry over to final execution? With the number of candidates, the selection pipeline, and the inference budget fixed, what do EF and Stage 3 each contribute?

A2. We add a public EF comparison and a public closed-loop evaluation after two epochs of Stage 3 adaptation, together with internal execution-level results.

To clarify the configuration first: the original public BASE model uses Stage 1 + Stage 2, without Stage 3. The new RL model adds 2 epochs of reward adaptation. All configurations below use 3,000 candidates and FM×2 and keep the scoring head and execution post-processing; when EF is enabled it is EF×2. The models before and after RL use the same number of candidates and the same EF/FM call budget. The no-EF comparison keeps the number of FM calls fixed, so its total compute is lower by the two EF calls.

ConfigurationVal14 NRVal14 RTest14-hard NRTest14-hard ROverall
BASE: EF + scoring head84.8981.3164.0165.9574.08
No EF, scoring head kept81.3479.1364.1364.3572.28
RL: EF + scoring head83.1181.1566.9066.8974.55

Common valid counts: 99/99/99/97 (394 total). Overall is weighted by these counts. The same exception tokens are excluded across configurations; completed zero scores are kept. Deltas use unrounded means.

EF adds 1.80 points to the pooled score. Two RL epochs change Test14-hard NR/R by +2.89/+0.94 points and Val14 NR/R by −1.79/−0.15 points, yielding +0.47 points overall. Sub-metric means on the reported scenarios (0–100, higher is better):

Sub-metricBASENo EFRLNo EF − BASERL − BASE
No at-fault collision92.7792.8994.16+0.13+1.40
Drivable-area compliance96.4596.1996.95−0.25+0.51
Making progress91.8890.8691.62−1.02−0.25
Progress along route67.0166.1467.39−0.88+0.37
Driving-direction compliance98.7398.6098.98−0.13+0.25
Time to collision (TTC)87.3187.3188.580.00+1.27
Speed-limit compliance99.2899.2399.28−0.050.00
Comfort91.1290.1091.37−1.02+0.25
AuthorsOpenReview ↗

Response to Reviewer UaLS (Part 2/2): Closed-Loop Analysis (cont.) and Deployment Statistics (W2/Q1, Q2)

The RL gains are concentrated in the safety-related items (no at-fault collision +1.40, TTC +1.27, drivable-area compliance +0.51), while mean route progress changes by +0.37 and making progress by −0.25; removing EF mainly lowers making progress, route progress, and comfort. These descriptive changes measure closed-loop behavior after scoring-based selection and execution and complement the pre-selection candidate metrics in the paper.

The internal closed-loop evaluation before and after RL covers a further 21,082 cases. The number of passing cases rises from 9,441 to 9,656, and the pass rate from 44.78% to 45.80% (+1.02 percentage points, a net gain of 215 cases). The pass rates for nudging and lane changing rise by 3.00 and 2.03 percentage points. This evidence measures system behavior after downstream selection and execution, and it includes deployment-relevant interactive scenarios.

The three kinds of evidence thus have distinct roles: the candidate metrics in the paper assess generation quality, the public EF/RL comparisons assess component effect on public tasks, and the internal pass rates assess execution-level benefit in practical scenarios.

Q2: Can the authors report full-pipeline latency, and the scale of real-vehicle operation with takeover and failure statistics?

A3. The timing in the paper refers specifically to the trajectory generation module; we also add the on-vehicle configuration and the operational context.

The 2.06 ms in the paper is the average generation time of EF×2 + FM×2 on NVIDIA Drive Orin in float16. It excludes the encoder and the selector. It measures the cost of the generation module that this paper optimizes. Stage 3 reward queries and parameter updates happen during offline training; at inference time the system uses the existing decoder and call budget, with no additional reward queries or online optimization module.

The learned scoring head is used only for the nuPlan evaluation. To balance trajectory diversity and the real-time compute budget, the on-vehicle system randomly samples six anchors from the vocabulary, generates six candidates, and uses a rule-based selector to choose the executed trajectory. The system is deployed on thousands of autonomous vehicles operating in urban areas across dozens of cities, with millions of kilometers of autonomous driving accumulated. The autonomous-driving share is 99%.

These figures describe the deployment context. Takeover, fallback, and failure rates involve company-confidential information that we are not able to disclose here. The quantitative execution-level benefit is given by the closed-loop evaluations above.

Thank you for the suggestions on component attribution and evaluation rigor. If these responses and the additional experiments address your main concerns, we hope you might consider raising your score, if possible. If any issue still affects your assessment, please let us know and we will respond further. Thank you for your time and your review.

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

DriveAnchor combines a trajectory vocabulary with flow-matching residual decoding, corridor-conditioned anchor correction and zeroth-order reward fine-tuning. The latter perturbs trajectories along nearby anchor directions and backpropagates black-box collision feedback through the decoder. Evaluation includes sampled nuPlan scenarios, candidate-level analysis on internal data and real-vehicle integration with on-device timing.

Critical strengths
  1. The deployment-oriented design includes real-vehicle integration and measured generation time on Orin. The anchor-guided zeroth-order update is a practical way to obtain black-box feedback for a deterministic decoder, and its smoothed-surrogate interpretation is stated with appropriate scope.

  2. The internal experiments show substantial candidate-level collision reductions and include useful analyses of energy-field calls, flow-matching calls and screening. The limitations of the fixed vocabulary and static-only reward are acknowledged.

Critical weaknesses
  1. The public evidence is too limited to establish the claimed closed-loop advantage. Only 100 sampled scenarios are evaluated per setting, several reproduced baselines are much weaker than their published results, and strong standard baselines are absent. The main component evidence is instead proprietary and measured before the downstream selector. Candidate-level static-collision gains do not by themselves establish safer executed plans. Matched public closed-loop ablations are needed to connect the components to the driving claim.

  2. The reward and collision statistics in Table 2 appear inconsistent under the stated reward definition. If collision-free candidates receive 81, the reported collision frequencies imply reward bounds that seem incompatible with the reported means. This could reflect different evaluation sets, horizons or conventions rather than an error, but it must be reconciled before the Stage-3 result can be interpreted. Comparisons with deterministic GRPO/DPO surrogates also leave the advantage over established reward-guided alternatives unclear. I currently lean toward weak rejection.

Questions for authors
  1. Please specify exactly which stages and selector are used for nuPlan. What are the full-split results and closed-loop changes from removing Stage 3 or EF under the same protocol? Why do the reproduced baselines differ substantially from published results?

  2. Are Table 2 reward means and short-/long-horizon collision rates computed on the same candidates and horizons? Please show how each reward mean follows from Eq. (12), including collision indexing and any normalization.

  3. How much of the candidate-level improvement survives downstream selection, including interactions with dynamic agents? A matched reward-weighted regression or standard zeroth-order control would help isolate the value of anchor-guided directions.

4 comments
AuthorsOpenReview ↗

Response to Reviewer oKre (Part 1/2): nuPlan Configuration, Closed-Loop Ablations, and Table 2 Reward Statistics (W1/Q1, W2/Q2)

We thank the reviewer for recognizing the real-vehicle integration, the anchor-guided black-box update, and the analysis of the locally smoothed objective.

W1 & Q1: Public evidence is limited; the nuPlan stage and selection configuration should be stated, with closed-loop comparisons for EF and Stage 3.

A1. The original public model did not go through Stage 3. We add an EF comparison and public closed-loop results after two epochs of reward adaptation.

The original public BASE model uses Stage 1 + Stage 2, 3,000 candidates, EF×2, FM×2, the nuPlan scoring head, and execution post-processing. The no-EF configuration sets the EF calls to 0 and keeps everything else; the RL configuration keeps the BASE candidate count and inference budget after two epochs of reward adaptation.

ConfigurationVal14 NRVal14 RTest14-hard NRTest14-hard ROverall
BASE: EF + scoring head84.8981.3164.0165.9574.08
No EF81.3479.1364.1364.3572.28
RL83.1181.1566.9066.8974.55

Common valid counts: 99/99/99/97 (394 total). Overall is weighted by these counts. The same exception tokens are excluded across configurations; completed zero scores are kept. Deltas use unrounded means.

EF adds 1.80 points to the pooled score. Two RL epochs change Test14-hard NR/R by +2.89/+0.94 points and Val14 NR/R by −1.79/−0.15 points, yielding +0.47 points overall. Sub-metric means on the reported scenarios (0–100, higher is better):

Sub-metricBASENo EFRLNo EF − BASERL − BASE
No at-fault collision92.7792.8994.16+0.13+1.40
Drivable-area compliance96.4596.1996.95−0.25+0.51
Making progress91.8890.8691.62−1.02−0.25
Progress along route67.0166.1467.39−0.88+0.37
Driving direction98.7398.6098.98−0.13+0.25
Time to collision (TTC)87.3187.3188.580.00+1.27
Speed limit99.2899.2399.28−0.050.00
Comfort91.1290.1091.37−1.02+0.25

The RL gains are concentrated in the safety-related items (collision +1.40, TTC +1.27, drivable area +0.51) while mean route progress changes by +0.37 and making progress by −0.25; on Test14-hard NR these three safety items are +2.53, +3.03, and +1.01. This matches the Stage 3 reward, which penalizes collisions with static obstacles and road boundaries. Removing EF mainly lowers making progress, route progress, and comfort. These are descriptive changes on the reported subsets.

Every method in Table 1 is evaluated on the same fixed list of 100 sampled scenarios per setting; several baselines use the reported "w/o refine" configuration. The scores describe these configurations and lists. The evidence has three complementary roles: (1) public EF/RL comparisons on matched scenarios assess component effects; (2) internal closed-loop evaluation covers 21,082 cases, including nudging, lane changing, merging, intersection turns, overtaking, yielding, and emergency braking. Passing cases rise from 9,441 to 9,656, pass rate from 44.78% to 45.80%, and nudging/lane-changing pass rates by 3.00/2.03 percentage points. Separate candidate analysis covers 10,000 scenes with 2,398 candidates each; (3) deployment spans thousands of autonomous vehicles in dozens of cities, with millions of kilometers of urban autonomous driving accumulated. The autonomous-driving share is 99%.

W2 & Q2 (reward statistics): Do the mean rewards in Table 2 conflict with the collision rates and Eq. (12)?

A2. All five means fall within the interval permitted by the corresponding collision rates.

Eq. (12) uses zero-based indexing. For T = 80, a colliding candidate receives the index of its first collision, j ∈ {0, …, 79}, and a collision-free candidate receives 81. A collision therefore does not imply a reward of zero, and the mean depends on when the first collision occurs. With p_S = P(R < 40) and p_L = P(R < 80):

81(1−pL)+40(pL−pS)≤E[R]≤81(1−pL)+79(pL−pS)+39pS.81(1-p_L)+40(p_L-p_S)\leq \mathbb E[R]\leq 81(1-p_L)+79(p_L-p_S)+39p_S.

ConfigurationAdmissible intervalMean in Table 2
FM×246.9306–68.981259.22
FMRL×276.8925–79.681078.46
EF×1 + FM×164.9773–75.222670.13
EF×1 + FMRL×174.4782–79.072477.14
EF×1 + FMRL×276.6687–80.073478.87

Every configuration in Table 2 is evaluated on the same main evaluation set and the full candidate pool (10,000 scenes with 2,398 candidates each), before residual filtering and final selection. The short and long horizons use indices 0–39 and 0–79; the mean is the direct average of the Eq. (12) reward, with no normalization.

AuthorsOpenReview ↗

Response to Reviewer oKre (Part 2/2): Execution-Level Benefit and Alternative Optimizers (Q3)

Q3 (execution-level benefit): How much of the candidate-level improvement survives downstream selection, and does it involve dynamic traffic?

A3. The new public comparisons and the internal closed-loop evaluation measure final system behavior directly.

The public closed-loop comparison and the internal closed-loop evaluation are reported in A1. Both are measured after downstream selection and execution, and internal categories such as nudging and lane changing involve other traffic participants.

To balance trajectory diversity and the real-time compute budget, the on-vehicle system randomly samples six anchors from the vocabulary, generates six candidates, and uses a rule-based selector to choose the executed trajectory. The system has accumulated millions of kilometers of urban autonomous driving across dozens of cities. The autonomous-driving share is 99%.

The suggestion on dynamic traffic is helpful. As next steps, we plan to condition vocabulary-neighborhood selection and perturbation ranges on ego initial speed, dynamics requirements, interaction cues, and traffic-rule preferences.

W2 & Q3 (alternative optimizers): Is the GRPO/DPO surrogate comparison sufficient? Is there a comparison with standard ZO or reward-weighted regression?

A4. The paper already includes a standard ZO comparison.

Appendix I.3 (Table 8) compares standard ZO with Anchor ZO on 1,000 held-out intersection scenes. Both share the Stage 2 initialization and use N = 16, 17 reward queries per center, and the same distance scaling rule:

MetricStandard ZOAnchor ZO
minADE@80 (m)1.781.50
Static collision rate, all candidates1.97%0.28%
Static collision rate, residual top-501.11%0.44%

This compares complete sampling configurations, in which both the direction distribution and the radial perturbation change, so it tests the configuration as a whole. A matched reward-weighted regression control is left for future work.

On the advantage over the deterministic GRPO/DPO surrogates: Appendix I.2 (Table 7) compares the three update rules on transfer between maneuver subsets over 10,000 evaluation scenarios (collision rate in %, mean ± SD over three repetitions):

Transfer directionZO (ours)GRPO-inspiredDPO-inspired
Jct→Lac, FM×14.41±0.606.65±0.967.04±0.61
Jct→Lac, FM×21.77±0.553.75±0.903.72±0.59
Lac→Jct, FM×13.49±0.687.07±0.905.02±1.05
Lac→Jct, FM×21.57±0.583.80±0.501.44±0.43

ZO has the lowest reported mean collision rate in three of the four settings. For Lac→Jct with FM×2, the DPO-inspired mean is 1.44%, compared with 1.57% for ZO. Each transfer test reward-fine-tunes the model on one maneuver subset and evaluates it on the other. These results provide evidence of transfer across the reported maneuver types.

The reward update supports efficient, extensible adaptation of an existing generator. Once a center trajectory is generated, exploration probes are constructed directly in trajectory space and scored without rerunning the generator for each probe. The vocabulary, neighbor-selection rule, and perturbation scale provide explicit interfaces for incorporating motion constraints and driving preferences. Engineers can shape the exploration neighborhood and replace the black-box reward while retaining the generator architecture and decoding procedure. Our GRPO/DPO-inspired controls compare surrogate updates for the same deterministic decoder.

Thank you for your attention to numerical consistency and execution-level benefit. If these responses and the additional experiments address your main concerns, we hope you might consider raising your score, if possible. If any key issue still affects your assessment, please let us know and we will respond to it specifically. Thank you for your time and your review.

AuthorsOpenReview ↗

A Brief Follow-up

Dear Reviewer oKre,

Apologies for this brief follow-up. We would be grateful if, when convenient, you could let us know whether our responses and additional results have addressed your main concerns. If any questions remain, we would be happy to provide further clarification. If our responses have resolved your concerns, we sincerely hope that you will consider giving the paper a higher score.

Thank you again for your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

Follow-up on closed-loop results and reward consistency

Dear Reviewer oKre,

As the author–reviewer discussion draws to a close, we would like to follow up on your questions. DriveAnchor is designed to help teams improve an existing planner when road tests reveal a new failure. Collecting corrective demonstrations takes time. Teams can steer the trained planner by adjusting the corridor supplied to EF, or use reward feedback on a small set of existing logs to fine-tune the decoder offline in Stage 3. The closed-loop ablations test whether these changes improve the plan that is ultimately executed.

  1. Performance after trajectory selection. On 393 matched nuPlan scenario–mode pairs, we fixed the candidate count, FM call count, score head, and downstream processing. The mean closed-loop score rose from 72.46 without EF to 74.27 with EF and 74.74 after Stage 3. On the 25 matched lane-changing episodes, EF raised the mean score by 4.89 points, and Stage 3 added a further 1.42 points.
  2. Diffusion Planner comparisons. Using the same 21,365 initial training frames, DriveAnchor scored higher than limited-data Diffusion Planner (DP) in all four settings. Starting from the official pretrained DP checkpoint, Stage 3 scored 78.79 versus 78.23 for supervised fine-tuning on the same 393 pairs, with adaptation data and update counts matched.
  3. Reward consistency. The mean reward and both collision rates in the original Table 2 (now Table 3) use the same candidate pool. The reward is the first collision index (0–79), or 81 for a collision-free trajectory, and is averaged directly. For the two-call FMRL configuration, the short- and long-horizon collision rates of 2.94% and 7.15% imply a mean reward between 76.89 and 79.68, consistent with the reported 78.46.
  4. Standard ZO control. Appendix I.3 compares the two sampling configurations using 17 reward queries per center. The candidate static collision rate was 0.28% for anchor-guided sampling and 1.97% for standard ZO. The comparison covers differences in both perturbation direction and the distribution of perturbation lengths.

If the new closed-loop results and reward analysis address your main concerns, we respectfully ask that you consider raising your rating.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.