acceptodds
Under review as a conference paper at ICLR 2027

RISE: Enhancing Safety for End-to-End Autonomous Driving through Risk-Aware Energy-Guided Diffusion Planner

Abstract

In end-to-end autonomous driving, vision-language-action (VLA) planners still tend to generate the high-risk trajectories in some long-tail scenarios. Existing safety-oriented approaches rely mostly on the post-training, which requires a costly fine-tuning of the pre-trained backbone and often degrades its general driving capability, thus leading to an inevitable safety-efficiency trade-off. In this paper, we propose RISE, a RISk-aware Energy-guided planning framework that incorporates these safety constraints into the trajectory generation of a diffusion planner, without any additional fine-tuning of the pre-trained VLA backbone. Specifically, we design a mixed energy guidance consisting of a rule-based guidance, an expert-based guidance and a risk assessor. The rule-based guidance constructs differentiable energy fields from drivable-area compliance and lane-keeping constraints, while the expert-based guidance takes reference trajectories from a safety-specialized expert trained separately via a safe reinforcement learning variant. Since the expert is inherently conservative, a lightweight rule-based risk assessor is further designed to activate the expert guidance only when potential risk is detected, thus preserving driving efficiency in ordinary scenarios. Experiments show that RISE improves the ReCogDrive baseline by 1.1 PDMS on NAVSIM (91.5) and by 24.6 PDMS on its high-risk subset, and achieves a Driving Score of 85.52 with a 59.09% Success Rate on the Bench2Drive closed-loop benchmark, outperforming the baseline by 12.03 and 17.27 percentage point, respectively.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-language-action model, autonomous driving, energy guidance, safe reinforcement learning

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_fZ5hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RISE, a safety-oriented framework for diffusion-based VLA planning. It combines differentiable road-rule energies with guidance toward a separately trained safety expert, activated by a heuristic risk assessor. The base VLA remains frozen, while the expert is trained using Safe-GRPO. Evaluation on NAVSIM and Bench2Drive reports improvements in planning scores and safety-related metrics.

Critical strengths
  • The framework addresses a practical safety-efficiency trade-off. Selectively applying expert guidance is a reasonable way to use a conservative expert without replacing the baseline policy in normal scenarios.
  • The reported Bench2Drive improvements are substantial: Driving Score increases from 73.49 to 85.52, and Success Rate from 41.82% to 59.09%. These closed-loop results suggest practical value beyond the smaller average gain on NAVSIM.
Critical weaknesses
  • The framework brings established ideas in rule-based energy guidance, backward guidance, constrained policy optimization, and risk gating. While the combination is reasonable, its methodological contribution could be clearer. In particular, the limitations of clean-trajectory guidance are not discussed. The energy gradients at a single denoised estimate generally differ from noisy-distribution guidance, especially at high noise levels. Also the comparisons with standard GRPO or fixed-penalty training would help establish the benefit of Safe-GRPO.
  • The results suggest practical value, although the contribution of each design choice remains somewhat unclear. On NAVSIM, the full method achieves 91.5 PDMS versus 91.3 for rule-based guidance alone. Comparisons with ungated combined guidance and direct risk-based expert switching would help explain this gain. The larger Bench2Drive improvements are encouraging, but component ablations would strengthen the evidence, especially given the latency increase from approximately 373 ms to 875 ms.
Questions for authors
  • Is backward guidance intended only as heuristic trajectory correction, or is there a distributional interpretation? Please clarify the approximation involved in evaluating energy on the predicted clean trajectory.
  • How is the high-risk NAVSIM subset defined, how large is it, and was its selection fixed before hyperparameter tuning? The large gain on this subset is central to the paper’s empirical argument.
  • Could you compare Safe-GRPO with standard GRPO or fixed-penalty training under the same settings? It would be helpful to compare against a simpler baseline that directly uses the expert’s trajectory when the same risk threshold is exceeded.
  • Could you clarify what RISE adds beyond existing energy-guided planners, such as GuideFlow, and which experimental findings specifically support this advance?
7 comments
AuthorsOpenReview ↗

Official Response by Authors to Reviewer fZ5h (1/7)

Dear Reviewer fZ5h,

We thank the reviewer for the constructive comments, and for recognizing selective expert guidance as a practical answer to the safety-efficiency trade-off and the closed-loop gains on Bench2Drive. We added guidance-timing, ungated, and direct-switching controls, seed variation, a Bench2Drive component ablation, and an optimized implementation, and will include them in the revised manuscript. In summary:

  • Clean-trajectory guidance (W1/Q1). Backward guidance has a distributional motivation (energy-reweighted sampling), but evaluating the energy at the point estimate x^0\hat{x}_0 is an approximation that ignores conditional variance. In line with the reviewer's concern, guidance at high noise does not help: early-only guidance gives 90.3 PDMS, vs. 91.4 for late-only and 91.5 for full RISE. We will add this limitation to the manuscript.
  • Gain over rule-only guidance (W2/Q3). The margin on full navtest is small but consistent across three seeds (91.48 ± 0.03 vs. 91.34 ± 0.01 PDMS). Both requested controls score lower: ungated combined guidance reaches 90.9 and direct expert switching 90.6. On Bench2Drive dev10, gated expert guidance raises DS from 74.40 to 79.70 at the same 60% SR.
  • Latency (W2). After optimization, RISE runs at 626.50 ms per plan (previously 874.58 ms; ReCogDrive 372.66 ms) with unchanged PDMS.
  • High-risk subset (Q2). The original 1,215-sample subset was fixed before tuning, and hyperparameters were selected by full-navtest PDMS. For a fairer evaluation, we will reselect the high-risk subset using only map and scene annotations, without using baseline predictions or scores. The revised paper will report results on this independently selected subset.
  • Safe-GRPO (W1/Q3). Compared with standard GRPO from the same IL checkpoint (the ReCogDrive RL stage), Safe-GRPO gives higher NC (98.7 vs. 97.7) and TTC (96.2 vs. 94.6) but lower EP (81.1 vs. 86.2).
  • Advances over GuideFlow (W1/Q4) are detailed below.

We hope these results address the reviewer's concerns, and we would be glad to clarify any remaining questions during the discussion period.

AuthorsOpenReview ↗

Official Response by Authors to Reviewer fZ5h (2/7)

Response 1 - Clean-trajectory guidance and its distributional approximation (W1 and Q1).

W1: The energy gradients at a single denoised estimate generally differ from noisy-distribution guidance, especially at high noise levels.

Q1: Is backward guidance intended only as heuristic trajectory correction, or is there a distributional interpretation? Please clarify the approximation involved in evaluating energy on the predicted clean trajectory.

Backward guidance has a distributional motivation but is implemented as an approximation. The target is energy-reweighted sampling, pE(x0∣c)∝pθ(x0∣c)exp⁡[−E(x0,c)]p_E(x_0\mid c)\propto p_\theta(x_0\mid c)\exp[-E(x_0,c)], whose exact correction averages the energy weight over pθ(x0∣xt,c)p_\theta(x_0\mid x_t,c). RISE instead evaluates the energy at the point estimate x^0(t)\hat{x}_0^{(t)}, which ignores conditional variance and multimodality, and this error is largest at high noise. Our timing ablation demonstrates that guidance in the first half of denoising alone does not help (90.3 vs. 90.4 PDMS for the baseline), while guidance in the second half alone gives 91.4, close to full RISE (91.5). We will add this approximation and its limitation to the paper.

Detailed responses:

What the point estimate approximates. Under Gaussian noising, a perfectly learned squared-error noise predictor gives x^0(t)=E[x0∣xt,c]\hat{x}_0^{(t)}=\mathbb{E}[x_0\mid x_t,c]. In practice, RISE uses the model's clean-trajectory prediction as a point estimate. The approximation is to evaluate the nonlinear energy weight at this point instead of averaging that weight over the conditional distribution. The guidance calculation therefore omits conditional variance and multimodality, which can matter especially at high noise.

Distributional interpretation and actual update. The distributional motivation is energy-reweighted generation, pE(x0∣c)∝pθ(x0∣c)exp⁡[−E(x0,c)]p_E(x_0\mid c)\propto p_\theta(x_0\mid c)\exp[-E(x_0,c)]. Its exact noisy-state score correction is ∇xtlog⁡ht(xt,c)\nabla_{x_t}\log h_t(x_t,c), where

ht(xt,c)=Epθ(x0∣xt,c)[exp⁡(−E(x0,c))].h_t(x_t,c)=\mathbb{E}_{p_\theta(x_0\mid x_t,c)}[\exp(-E(x_0,c))].

The point-estimate surrogate is ht(xt,c)≈exp⁡[−E(x^0(t),c)]h_t(x_t,c)\approx\exp[-E(\hat{x}_0^{(t)},c)], analogous to the point-estimate approximation in diffusion posterior sampling. RISE takes a few gradient steps on E(x^0(t)+Δ,c)E(\hat{x}_0^{(t)}+\Delta,c) and maps Δ\Delta to the noise prediction through Eq. (3). This is a distributional motivation; the clean-space update is an approximation, not the exact noisy-state score correction.

Guidance timing. Early-only and late-only guidance use the first and last halves of denoising, respectively. Last-step-only guidance applies the same backward update only at the final denoising step.

Table 1. Guidance timing on the full NAVSIM navtest split. All variants use the same guidance formulation with different active denoising steps.

MethodNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
ReCogDrive (Baseline)97.797.794.6100.086.290.4
Early-only guidance97.797.694.5100.086.190.3
Last-step-only guidance97.798.794.6100.086.891.1
Late-only guidance97.898.894.7100.087.191.4
RISE97.899.094.8100.087.191.5

Last-step-only guidance reaches 91.1 PDMS, compared with 91.4 for late-only, 91.5 for full RISE, and 90.3 for early-only guidance. These results show that high-noise guidance can be unreliable because the point-estimate surrogate omits the conditional variance and multimodality of pθ(x0∣xt,c)p_\theta(x_0\mid x_t,c) in ht(xt,c)h_t(x_t,c), an approximation that matters most at high noise. At low noise, even last-step-only guidance reaches 91.1 PDMS and late-only guidance reaches 91.4 PDMS, close to full RISE at 91.5, so low-noise guidance alone yields a considerable gain and favors late-step guidance in this configuration.

AuthorsOpenReview ↗

Official Response by Authors to Reviewer fZ5h (3/7)

Response 2 - Component contributions (W2).

W2: Comparisons with ungated combined guidance and direct risk-based expert switching would help explain this gain.

Q3: It would be helpful to compare against a simpler baseline that directly uses the expert’s trajectory when the same risk threshold is exceeded.

The +0.2 PDMS margin of RISE over rule-only guidance on full navtest is small but consistent across three seeds (91.48 ± 0.03 vs. 91.34 ± 0.01), and both requested controls score below both methods: ungated combined guidance reaches 90.9 and direct expert switching 90.6. On Bench2Drive dev10, gated expert guidance raises DS from 74.40 to 79.70 at the same 60% SR, and removing the gate lowers SR to 30%.

Detailed responses:

Always-on mixed guidance on navtest. Table 1. Component ablations on the full NAVSIM navtest split. Parentheses show the PDMS drop in points relative to full RISE (ID 5).

IDRule-BasedExpert-BasedRisk AssessorNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
1✗✗✗97.797.794.6100.086.290.4 (1.1↓)
2✓✗✗97.798.994.5100.087.191.3 (0.2↓)
3✗✓✗98.097.995.1100.086.090.7 (0.8↓)
4✓✓✗98.598.595.8100.084.990.9 (0.6↓)
5✓✓✓97.899.094.8100.087.191.5

Always-on mixed guidance achieves the highest NC (98.5) and TTC (95.8), but at the cost of EP. And it reaches a PDMS of 90.9 versus the full RISE(91.5).

Comparison with simpler alternative on NAVSIM. We compared our method with simpler alternatives that direct expert switching on the full navtest split. It keeps rule guidance enabled and outputs the safety expert’s trajectory when risk exceeds the threshold.

Table 2. Comparison with simpler alternatives on the full NAVSIM navtest split. Parentheses show the PDMS drop in points relative to full RISE.

MethodNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
ReCogDrive (Baseline)97.797.794.6100.086.290.4 (1.1↓)
Direct expert switching97.898.294.9100.085.990.6 (0.9↓)
Rule-only97.798.994.5100.087.191.3 (0.2↓)
RISE97.899.094.8100.087.191.5

Keeping rule guidance active and switching directly to the expert at high risk yields 90.6 PDMS, lower than rule-only guidance(91.3 PDMS), and that directly adopting the expert trajectory is inferior to using it as a soft reference. A direct switch takes the expert trajectory as is. Full RISE uses it as a reference during denoising while continuing to apply road-rule guidance. This gives the planner room to adjust for road compliance and preserve progress. Full RISE reaches 91.5 PDMS, with higher DAC and EP than these simpler alternatives.

W2: The larger Bench2Drive improvements are encouraging, but component ablations would strengthen the evidence

We evaluated the guidance components on the dev10 subsplit on Bench2Drive.

Table 3. Bench2Drive dev10 component ablation. A check mark indicates an enabled component.

RuleExpertGateDS ↑SR (%) ↑Eff ↑
✗✗✗67.0050.00107.00
✓✗✗74.4060.00111.22
✗✓✓81.0040.00113.76
✓✓✗79.2030.00116.90
✓✓✓79.7060.00124.41

Rule-only guidance raises SR from 50% to 60%. Adding gated expert guidance improves DS from 74.40 to 79.70 and efficiency from 111.22 to 124.41 while maintaining SR. Without gating, DS remains similar (79.20), but SR drops to 30%. Gated expert guidance alone achieves the highest DS (81.00), but full RISE achieves higher SR (60% vs. 40%) and efficiency (124.41 vs. 113.76), providing a better balance across these metrics.

Reviewer_s5agOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper improves a frozen VLA diffusion planner by introducing energy guidance at inference time. The method combines a rule-based penalty mechanism with trajectory guidance generated by a separately trained safety expert. Additionally, a heuristic risk evaluator is designed to selectively enable expert guidance, thereby reducing unnecessary conservatism in common scenarios.

Critical strengths
  1. Well-motivated. Separating the original planner from safety-oriented interventions is a practical approach. Combining explicit road constraints with trained experts provides complementary sources of guidance, while risk-based interventions address the tendency of experts to act conservatively. Although this framework requires the separate training of safety experts, the original planning model remains unchanged.
  2. Encouraging closed-loop results. On Bench2Drive, Driving Score increases from 73.49 to 85.52, and Success Rate increases from 41.82% to 59.09%.
Critical weaknesses
  1. The concept is generally rather uninteresting, and the effectiveness is exaggerated.
  2. The 24.6-point improvement achieved by PDMS in high-risk scenarios is a key result, but the paper does not explicitly specify the size of this subset, the selection criteria, or the scenario identifiers. Furthermore, it is unclear whether this subset was determined independently of the methods being evaluated.
  3. The NAVSIM guidance scheme uses real-world road and lane information, whereas the closed-loop implementation uses YOLOv2 as these inputs. The paper should clarify whether the reproduced baseline models received equivalent structured information.
  4. Lack of Necessary Ablation.
  5. Limited visualization of the guidance and risk assessment processes makes it difficult to understand how risk scores evolve, when expert guidance is activated, and how each guidance component shapes intermediate trajectories during denoising.
Questions for authors

Refer to Weaknesses.

6 comments
AuthorsOpenReview ↗

Official Response by Authors to Reviewer s5ag (1/6)

Dear Reviewer s5ag,

We thank the reviewer for the comments, and for recognizing the motivation of keeping the planner frozen while adding complementary safety guidance, and the closed-loop gains on Bench2Drive. We added component ablations on NAVSIM and Bench2Drive and simpler-alternative controls, and will include them in the revised manuscript. In summary:

  • Contribution (W1). The gain comes from how the expert is used, not from adding a second energy term. With the same expert and maps, always-on mixed guidance (90.9 PDMS) and direct expert switching (90.6) both fall below rule-only guidance (91.3), while risk-gated reference guidance rises above it (91.5). On Bench2Drive dev10, removing the gate lowers SR from 60% to 30%.
  • Overstated claims (W1). We agree that several statements were too strong. We will replace "training-free" with "without fine-tuning the base planner", clarify the high-risk subset selection criteria, and separate the open-loop NAVSIM and closed-loop Bench2Drive claims.
  • High-risk subset (W2). The original subset contains 1,215 navtest samples: 607 baseline failures on NC, DAC, or TTC, and 608 samples selected from map and scene annotations only. For a fairer evaluation, we will reselect the high-risk subset using only map and scene annotations, without using baseline predictions or scores. The revised paper will report results on this independently selected subset.
  • Structured inputs (W3). The reproduced baseline does not receive road information, whereas RISE uses ground-truth maps on NAVSIM and YOLOPv2 predictions on Bench2Drive. The rule-only variant receives the same road input, and RISE still improves over it (91.48 ± 0.03 vs. 91.34 ± 0.01 PDMS over three seeds; 79.70 vs. 74.40 DS on dev10).
  • Ablations (W4) and visualization of risk evolution and guidance (W5) are detailed below.

We hope these results address the reviewer's concerns, and we would be glad to clarify any remaining questions during the discussion period.

AuthorsOpenReview ↗

Official Response by Authors to Reviewer s5ag (2/6)

Response 1 - Methodological contribution and evidence of effectiveness (W1).

W1: The concept is generally rather uninteresting, and the effectiveness is exaggerated.

The gain of RISE comes from how the expert is used, not from adding a second guidance term. With the same expert and maps, adding the expert energy at every step (90.9 PDMS) or outputting the expert trajectory at high risk (90.6) both score below rule-only guidance (91.3), whereas risk-gated reference guidance scores above it (91.5). In addtion, on Bench2Drive dev10, removing the gate lowers Success Rate from 60% to 30%. The rule-based gate matches a learned risk predictor without training one (91.5 vs. 91.3 PDMS, Appendix C.7).

We agree that several statements in the submission are too strong, and will revise them: "training-free" will become "without fine-tuning the base planner", with the expert's training cost reported. We will clarify how the high-risk subset was selected (Response 2). The conclusion will distinguish open-loop NAVSIM from closed-loop Bench2Drive. And the DS gain will be reported in points rather than percentage points.

Detailed responses:

1. The simple ungated combination does not preserve the safety-progress balance. Compared with a simple combination of a rule-based penalty mechanism with trajectory guidance(a always-on mixed guidance without the risk gate), RISE keeps rule-based guidance active and applies the safety expert's reference energy only when the current trajectory estimate exceeds the risk threshold. On full navtest, always-on mixed guidance reaches a PDMS of only 90.9. Risk gating recovers progress and gives the highest PDMS among these variants. On Bench2Drive dev10, gating improves SR(30 -> 60%) and efficiency(116.90 -> 124.41)(More analysis of the ablation results is detailed in Response 4). The benefit is a better balance across these measures. Thus the gain comes from selective expert-based guidance, not from mechanically adding two guidance terms.

Table 1. Bench2Drive dev10 component ablation.

RuleExpertGateDS ↑SR (%) ↑Eff ↑
✗✗✗67.0050.00107.00
✓✗✗74.4060.00111.22
✗✓✓81.0040.00113.76
✓✓✗79.2030.00116.90
✓✓✓79.7060.00124.41

Table 2. Component ablations on the full NAVSIM navtest split.

IDRule-BasedExpert-BasedRisk AssessorNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
1✗✗✗97.797.794.6100.086.290.4 (1.1↓)
2✓✗✗97.798.994.5100.087.191.3 (0.2↓)
3✗✓✗98.097.995.1100.086.090.7 (0.8↓)
4✓✓✗98.598.595.8100.084.990.9 (0.6↓)
5✓✓✓97.899.094.8100.087.191.5

2. Trajectory-dependent risk checks without an extra network. A learned risk predictor must generalize to intermediate trajectories that change throughout denoising. Our assessor directly checks each candidate against predicted obstacle motion and road constraints, making intervention decisions depend on the generated motion without training an additional network. The resulting planner reaches 91.5 PDMS versus 91.3 with the learned assessor (Appendix C.7, Table 7).

3. Each addition addresses an observed limitation. Diffusion Planner uses hand-designed energies for collision avoidance, road compliance, and comfort. These provide explicit control, but covering diverse interactions requires additional manual energy design, making comprehensive coverage difficult.

ReCogDrive uses RL post-training to update the planner's policy. Stronger safety optimization may also make ordinary driving more conservative, sacrificing progress even where additional caution is unnecessary. This trade-off is visible in our own safety-focused post-training experiment: the expert improves NC and TTC but lowers EP from 86.2 to 81.1 and PDMS from 90.4 to 89.5.

RISE therefore keeps the original planner frozen, integrates explicit rule constraints with a learned trajectory reference, and gates its application by risk. It reaches 91.5 PDMS on NAVSIM, versus 89.5 for the standalone expert. On Bench2Drive dev10, it raises rule-only DS from 74.40 to 79.70 and efficiency from 111.22 to 124.41, at the same 60% SR.

AuthorsOpenReview ↗

Official Response by Authors to Reviewer s5ag (3/6)

Response 2 - High-risk subset size, selection criteria, and independence (W2).

W2: The 24.6-point improvement achieved by PDMS in high-risk scenarios is a key result, but the paper does not explicitly specify the size of this subset, the selection criteria, or the scenario identifiers. Furthermore, it is unclear whether this subset was determined independently of the methods being evaluated.

The subset contains 1,215 samples, about 10% of navtest. We selected 607 samples from the low-scoring ReCogDrive cases using NC, DAC, and TTC, which is independent of our methods, but depends on our baseline. We selected the other 608 from map and scene annotations using lane curvature, vehicle and pedestrian density, intersections, and the presence of barriers. The first group measures recovery from baseline failures. The second group does not depend on any planner’s predictions.

Reviewer_VrjROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RISE adds risk-aware energy guidance to a ReCogDrive diffusion planner. Rule-based map terms guide DDIM sampling, while proximity to a separately trained safety expert's trajectory is activated above a risk threshold. The paper reports improved NAVSIM scores, a large gain on a high-risk subset and substantial Bench2Drive improvements, with latency and failure-case analyses.

Critical strengths
  1. Long-tail safety is an important problem, and risk-gated expert guidance offers an intuitive way to balance caution against progress. The expert-only ablation makes this trade-off visible rather than treating a more conservative policy as uniformly better.

  2. The reported Bench2Drive improvements are substantial and give the method potentially meaningful closed-loop value. Reporting latency and failure cases also helps assess its practical cost.

Critical weaknesses
  1. Input and compute differences confound the comparisons. NAVSIM rule guidance uses ground-truth drivable-area and lane information, unlike the camera-only baseline. The method also trains a separate safety expert, and reported inference latency increases from about 373 to 875 ms. Freezing the base planner therefore does not make the full system training-free or cost-free. Matched-input and matched-compute controls are needed to attribute the gain to the proposed guidance.

  2. The headline safety evidence and guidance design need clearer validation. The definition and size of the high-risk subset are not clear, nor is the split used to select thresholds and guidance weights. Small aggregate margins have no reported seed variation. A hard switch to the safety expert and risk-gated rule-only guidance would test whether iterative energy guidance adds value beyond simpler alternatives. The unclear NAVSIM-to-CARLA expert protocol also limits interpretation of the strong closed-loop result. I currently lean toward weak rejection.

Questions for authors
  1. What are the NAVSIM results using predicted maps or an input-matched baseline? Please report the complete safety-expert training cost and compare with an equivalent post-training or simple expert-switching baseline.

  2. How is the high-risk subset defined, how many scenarios does it contain, and was the definition fixed independently of the evaluated methods' outcomes? Which split selected thresholds and guidance weights? Please distinguish descriptive navtest sweeps from actual model selection.

  3. Was the expert retrained for CARLA or transferred from NAVSIM? How was ReCogDrive reproduced, and how do map-estimation errors affect results? A hard expert switch and risk-gated rule-only control would clarify the need for the full guidance mechanism.

5 comments
AuthorsOpenReview ↗

Official Response by Authors to Reviewer VrjR (1/5)

Dear Reviewer VrjR,

We thank the reviewer for the detailed comments, and for recognizing the risk-gated design, the expert-only ablation that exposes the safety-efficiency trade-off, and the closed-loop gains on Bench2Drive. We added seed variation, simpler-alternative controls, a Bench2Drive component ablation, and an optimized implementation, and will include them in the revised paper. In summary:

  • Map input (W1/Q1). We agree that the NAVSIM inputs are not matched: RISE computes rule energy from ground-truth maps. The rule-only variant receives the same maps, so the margin of RISE over it measures what the expert guidance adds beyond the map input: 91.48 ± 0.03 vs. 91.34 ± 0.01 PDMS over three seeds on full navtest, and 79.70 vs. 74.40 DS on Bench2Drive dev10, where roads are predicted by YOLOPv2 instead of taken from simulator maps.
  • Compute and training cost (W1/Q1). After optimization, RISE adds 253.84 ms per plan instead of 501.92 ms (626.50 vs. 372.66 ms for ReCogDrive) with unchanged PDMS. The safety expert costs 120 GPU-hours, the same training time as ReCogDrive's RL stage. Deploying this safety post-trained model alone gives 89.5 PDMS, vs. 91.5 for RISE. We will replace "training-free" with "without fine-tuning the base planner".
  • Simpler alternatives (W2/Q1/Q3). With the same expert, threshold, and maps, outputting the expert trajectory at high risk lowers PDMS from 91.3 (rule-only) to 90.6, whereas using it as a denoising reference raises PDMS to 91.5. Risk-gated rule-only guidance gives no gain over the baseline (90.4).
  • High-risk subset and model selection (W2/Q2). Of the 1,215 samples in the original subset, 607 are baseline failures and 608 are selected from scene annotations only. For a fairer evaluation, we will reselect the high-risk subset using only map and scene annotations, without using baseline predictions or scores. The revised paper will report results on this independently selected subset. Final hyperparameters were selected on navtest, and Bench2Drive reuses them without re-tuning.
  • CARLA expert, ReCogDrive reproduction, and map-estimation errors (Q3) are detailed below.

We hope these results address the reviewer's concerns, and we would be glad to clarify any remaining questions during the discussion period.

AuthorsOpenReview ↗

Official Response by Authors to Reviewer VrjR (2/5)

Response 1 - Map input fairness (W1 and Q1).

W1: NAVSIM rule guidance uses ground-truth drivable-area and lane information, unlike the camera-only baseline.

W1: Matched-input and matched-compute controls are needed to attribute the gain to the proposed guidance.

Q1: What are the NAVSIM results using predicted maps or an input-matched baseline?

We agree that the NAVSIM inputs are not matched: RISE computes rule energy from ground-truth maps, which ReCogDrive does not receive, so part of the NAVSIM gain over the baseline may come from this input. The rule-only variant receives the same maps, so the margin of RISE over rule-only guidance isolates what the expert guidance adds beyond the map input (91.48 ± 0.03 vs. 91.34 ± 0.01 PDMS over three seeds). On Bench2Drive, rule guidance uses YOLOPv2 road predictions from the camera images instead of simulator maps, and RISE improves DS/SR from 73.49/41.82% to 85.52/59.09% over the same reproduced checkpoint (220 routes).

Detailed responses:

NAVSIM v1 scores the entire four-second trajectory from a single prediction. During evaluation, the planner receives no new observations to update the later waypoints. Images taken at planning time may occluded lane boundaries, making the rule energy less reliable for the later part of the trajectory. For this reason, we compute rule energy from ground-truth maps in NAVSIM experiments.

On Bench2Drive, RISE and the reproduced ReCogDrive use the same camera inputs and closed-loop protocol, and neither uses the simulator's road annotations. RISE additionally runs YOLOPv2 to predict road structure. At each iteration, the vehicle executes a short prefix of the predicted trajectory and replans from the next observation. These repeated corrections can prevent errors from occluded lane boundaries in the distance, which helps explain why the benefits of RISE are more visible in closed loop.

W1: reported inference latency increases from about 373 to 875 ms.

We reduced the inference latency of RISE from 874.58ms (+134.69%) to 626.50ms (+68.12%) without any performance degradation.

Detailed responses:

We optimized computation using parallel processing, caching, and vectorized projections. We remeasured average per-plan latency on the same hardware and evaluation setup.

Table 1. Average per-plan inference latency before and after optimization.

Rule-BasedExpert-BasedRisk AssessorLatency before (ms)Overhead beforeLatency after (ms)Overhead after
✗✗✗372.66---372.66---
✓✗✗824.76+121.32%556.80+49.41%
✗✓✓689.48+85.02%474.50+27.33%
✓✓✓874.58+134.69%626.50+68.12%

Relative to ReCogDrive's 372.66 ms baseline, RISE's additional latency decreases by 49.4%, from 501.92 to 253.84 ms. Full navtest PDMS remains 91.5 before and after optimization. The Bench2Drive ablation also uses the optimized guidance computation.

AuthorsOpenReview ↗

Official Response by Authors to Reviewer VrjR (3/5)

Response 2 - Computational cost (W1 and Q1).

W1: Freezing the base planner therefore does not make the full system training-free or cost-free.

We agree, and will replace "training-free" with "without fine-tuning the base planner" and report the expert's training cost alongside the main results. RISE keeps the pretrained VLA and the diffusion planning head frozen and changes only the sampling updates at inference; all additional training is confined to the separate safety expert, whose cost is reported below.

Q1: Please report the complete safety-expert training cost

Training the NAVSIM expert takes 120 GPU-hours (15 hours on eight H800 GPUs), the same wall-clock time as our reproduction of ReCogDrive's third training stage, and the Bench2Drive expert takes 44 GPU-hours (5.5 hours). Using this budget for safety post-training alone, i.e., deploying the expert itself, gives 89.5 PDMS with EP dropping to 81.1, whereas using the same expert as gated guidance gives 91.5. The Bench2Drive expert is initialized from the two-stage checkpoint shared with our baseline reproduction, which we therefore do not count, and the risk assessor is rule-based and needs no training.

Q1: compare with an equivalent post-training or simple expert-switching baseline.

W2: A hard switch to the safety expert and risk-gated rule-only guidance would test whether iterative energy guidance adds value beyond simpler alternatives.

Both simpler alternatives score below RISE on full navtest: risk-gated rule-only guidance gives no gain over the baseline (90.4 PDMS), and direct expert switching reaches 90.6, vs. 91.5 for RISE. Direct switching uses the same expert, threshold, and rule guidance as RISE, but outputs the expert trajectory as is when the risk exceeds the threshold; this lowers PDMS below rule-only guidance (91.3), whereas using the same trajectory as a denoising reference raises it. The expert trajectory is therefore more useful as a reference for the planner to refine than as a replacement.

Detailed responses:

Table 2. Comparison with simpler alternative on the full NAVSIM navtest split.

MethodNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
ReCogDrive (Baseline)97.797.794.6100.086.290.4 (1.1↓)
Risk-gated rule guidance97.897.794.7100.086.290.4 (1.1↓)
Direct expert switching97.898.294.9100.085.990.6 (0.9↓)
Rule-only97.798.994.5100.087.191.3 (0.2↓)
RISE97.899.094.8100.087.191.5

Risk-gated rule guidance yields 90.4 PDMS, close to ReCogDrive’s 90.4. It applies rule guidance to only a small fraction of scenarios, and the hand-designed energies have limited coverage of interactions with other road users in high risk scenes. This helps explain the small gain on the full test set.

Keeping rule guidance active and switching directly to the expert at high risk yields 90.6 PDMS, lower than rule-only guidance(91.3 PDMS), and that directly adopting the expert trajectory is inferior to using it as a soft reference. A direct switch takes the expert trajectory as is, while full RISE uses it as a reference during denoising. This gives the planner room to adjust for road compliance and preserve progress.

Full RISE reaches 91.5 PDMS, with higher DAC and EP than these simpler alternatives.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their thoughtful evaluations and constructive feedback. The reviewers recognized long-tail safety and the safety–efficiency trade-off as important and practical problems (VrjR, fZ5h). The complementarity of explicit road constraints and trained experts, together with keeping the original planner unchanged, was considered well motivated (s5ag). Selective expert intervention was recognized as a reasonable way to balance caution against progress (VrjR, fZ5h). All reviewers highlighted the encouraging closed-loop results on Bench2Drive, with VrjR and fZ5h further recognizing their potential practical value.

The main concerns are about what RISE adds beyond existing guidance methods, how each component contributes, and how to interpret the evaluation and computational cost. We have supplemented the experiments and clarifications and revised the manuscript accordingly. The main updates are:

  • Expert-based energy guidance beyond simpler alternatives (VrjR, s5ag, fZ5h): RISE's contribution is converting a safety expert's scene-conditioned trajectory into a differentiable energy reference and using trajectory-dependent risk gating to selectively refine high-risk trajectory estimates during a frozen planner's sampling. Added full-navtest direct-expert-switching and risk-gated rule-only controls support this design. We also clarified the differences from GuideFlow and added a theoretical analysis of noise-dependent bias in point-estimate guidance. (Revised manuscript: Sections 3.2 and 5.3; Appendix D.2–D.3.)

  • Close-loop component ablation (VrjR, s5ag, fZ5h): On NAVSIM, we added an always-on mixed-guidance control to show how risk gating limits the progress loss from uniform expert intervention. On Bench2Drive, we added a complete set of component ablations to examine the contributions and complementarity of each component in closed-loop driving. (Revised manuscript: Section 5.3.)

  • Consistency and independent high-risk evaluation (VrjR, s5ag, fZ5h): Seed-variation experiments support the consistency of the full-navtest gains. We also updated high-risk results using a subset selected solely from map and scene annotations and clarified benchmark inputs and hyperparameter selection. (Revised manuscript: Sections 5.1 and 5.3.)

  • Lower inference latency overhead and explicit training costs (VrjR, fZ5h): We optimized guidance computation while preserving full-navtest performance and documented expert training costs and protocols. The responses also clarify CARLA expert retraining, baseline reproduction, and Safe-GRPO's adaptive penalty. (Revised manuscript: Section 5.2; Appendix A.)

  • Deeper theoretical analysis and guidance-timing experiments (fZ5h): We added a deeper analysis of noise-dependent bias in point-estimate guidance, explaining how the point-estimate approximation omits conditional variance and multimodality in energy-reweighted sampling. Early-only, late-only, and last-step-only ablations highlight the limited effectiveness of early, high-noise guidance, consistent with this approximation analysis. (Revised manuscript: Section 3.2; Appendix D.2.)

  • Visualization (s5ag): Added risk-score evolution and component-wise trajectory visualizations show how guidance changes the sampled trajectory during denoising and how rule constraints and learned expert guidance complement each other in scenarios where either alone is insufficient. (Revised manuscript: Appendix E.2–E.3.)

Point-by-point responses are provided under each review. We thank all the reviewers again for helping us improve the clarity and experimental validation of RISE.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.