RiSE-Drive: Risk-Sensitive Long-Horizon Candidate Evaluation for End-to-End Autonomous Driving

arXiv:2611.13790 · cs.RO · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "RiSE-Drive: Risk-Sensitive Long-Horizon Candidate Evaluation for End-to-End Autonomous Driving".

Title and authors: Rosa: The paper is titled RiSE-Drive: Risk-Sensitive Long-Horizon Candidate Evaluation for End-to-End Autonomous Driving, and it’s written by Yu Zhao, Hongyi Duan, Ming Yin, Hongyuan Zhang, Fan Yang, Hongzhi Wang, and Xiubo Liang.

Dev: That's a pretty big team behind this work in the robotics and autonomy space. It tells us this isn't just one person's idea; it’s a collaborative effort from different angles of the problem.

Rosa: Right. And what they’re doing is taking that candidate generation process—where you have a set of potential plans—and making it risk-sensitive over a long horizon, which is what makes this RiSE-Drive approach interesting for us to look at.

Dev: So, the core idea seems to be shifting from a deterministic oracle style evaluation, where you get one fixed score, to something that looks at the range of possible futures.

Rosa: Right. They first perform what they call candidate-conditioned latent future rollout—basically looking ahead based on which candidate plan you chose—then they estimate long-horizon outcome distributions instead of just a single scalar score.

Dev: That sounds like they’re trying to capture the uncertainty inherent in driving situations that we usually ignore when we're building deterministic models.

The paper's summary: Rosa: The summary says RiSE-Drive first does candidate-conditioned latent future rollout, then estimates long-horizon outcome distributions instead of single scalar scores, and finally ranks candidates using a tail-risk aware and perturbation robust decision criterion.

Dev: They’re essentially saying that the main difficulty isn't just producing plans that look plausible right now, but deciding which candidate is safest and most decision optimal under safety, progress, and rule compliance trade-offs.

Rosa: That’s a key point they make on page three of the work; they are focusing on deciding which candidate is safest and most decision optimal under those trade-offs.

Dev: They achieve this by predicting a compact set of future condition branches for each scene, performing candidate conditioned latent consequence propagation, decoding one long-horizon outcome sample for each branch, and then estimating an outcome distribution through ordered quantiles.

Rosa: And then they rank candidates using a criterion that combines the mean, lower tail Conditional Value at Risk with lambda, and cross condition instability. That’s a specific way they handle the risk part of things.

Dev: It sounds like they are explicitly trying to model those hidden tail failures that traditional point scores miss, by focusing on those lower quantiles where things go wrong.

The paper's improvements: Rosa: The paper suggests several improvements over older methods. First, they move away from just deterministic point scores and instead use distributional outcome assessment with explicit risk control.

Dev: That means the system is encouraged to prefer candidates that are not only strong under nominal conditions, but also reliable under long-horizon rollout and safety critical deviations.

Rosa: That’s the big implication there; they want candidates that are robust when things go wrong in a real driving situation, not just when everything is perfect.

Dev: They use a specific mathematical definition for the long horizon outcome sample, which they call y i l l l equals sum h from one to H of r i l h plus gamma H r term i l.

Rosa: That set y i l provides a finite sample representation of its long-horizon outcome distribution by pooling each branch specific latent consequence trajectory over time and then pooling across branches.

Dev: So they’re building this distribution head that outputs M ordered quantiles, q i equals h q (i), which is how they get their measure of the outcomes.

Conclusion: Rosa: So to wrap up, RiSE-Drive moves us toward risk sensitive candidate selection based on distributional outcomes, allowing us to rank candidates by mean plus lower tail conditional value at risk and instability.

Dev: It means we can explicitly model those hidden tail failures through ordered quantiles, which is something traditional point scores simply can't do when you need reliability under long horizon rollout.

Rosa: The experimental results on NAVSIM v1, NAVSIM v2, and HUGSIM show state-of-the-art performance, improving PDMS from ninety-three point one to ninety-four point two under matched DrivoR setup and raising the HD score on hard subsets significantly.

Dev: That gain is concentrated on the hard and extreme subsets where the HD score rises from eighteen point eight to twenty-two point seven and from thirty-two point five to thirty-nine which shows it really helps in those tough scenarios that matter most for safety cases <ref:2611.13790#pg3>.

Taro: From a failure mode perspective, I'm interested in how they handle the world misbehaving; the paper focuses on robustness under perturbation rather than just open-loop trajectory closeness.

Rosa: That’s right, and what I want to point out is that this framework learns a criterion via pairwise ranking loss, which favors trajectories that remain reliable not only on average but also under plausible downstream deviations and interaction heavy rollouts.

Dev: So the paper shows how we can use these distributional estimates to build a selector that prioritizes safety over just nominal speed, and it’s a solid framework for moving toward more reliable candidate generation.

Yu Zhao, Hongyi Duan, Ming Yin, Hongyuan Zhang, Fan Yang, Hongzhi Wang, Xiubo Liang

Zhejiang University

cs.RO

Submitted: 2026-11-10

Updated: 2026-11-10

Comments: 11 pages, 3 figures, 6 tables, including supplementary material. Published in ACM Multimedia 2026 (MM '26)

Journal ref: Proceedings of the 34th ACM International Conference on Multimedia (MM '26), 15482-15491 (2026)

DOI: 10.1145/3767308.3836337

Code: https://github.com/OpenDriveLab/OpenScene

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 98/100

The gist: RiSE-Drive presents a framework for risk-sensitive long-horizon candidate evaluation in end-to-end autonomous driving by reformulating candidate scoring from deterministic point scores into…

Key concepts

Distributional Outcome Assessment
Instead of giving a single score, the system estimates the full range of potential long-term results for each candidate across different future scenarios. This captures not just what is expected to happen, but also the probability and spread of various possible outcomes over time.
Long-Horizon Outcome Sample
This is a mathematical representation of a candidate's total performance over an extended period. It combines immediate rewards with future consequences propagated through different potential paths, providing a comprehensive view of the trajectory's ultimate result.
Risk-Sensitive Ranking Criterion
The final ranking score uses statistics like the mean and Conditional Value-at-Risk (CVaR) to explicitly prioritize safety. CVaR focuses on minimizing the impact of the worst outcomes, ensuring that candidates are preferred not just for high average performance but also for low tail risk.

Terminology

Summary

RiSE-Drive presents a framework for risk-sensitive long-horizon candidate evaluation in end-to-end autonomous driving by reformulating candidate scoring from deterministic point scores into distributional outcome assessment with explicit risk control The evaluator is encouraged to prefer candidates that are not only strong under nominal conditions, but also reliable under long-horizon rollout and safety-critical deviations

Core Problem Identification

Existing methods still mainly use deterministic point scores, heuristic subscores, or single-outcome evaluations, which makes it difficult to capture delayed failures, tail risk, and ranking instability under plausible future perturbations Candidate quality is not a deterministic scalar but a distribution of long-horizon consequences induced by uncertain future conditions Hidden tail failures remain invisible to the selector, and ranking becomes unstable exactly when robustness matters most

RiSE-Drive Methodology

RiSE-Drive first performs candidate-conditioned latent future rollout, then estimates long-horizon outcome distributions instead of single scalar scores, and finally ranks candidates using a tail-risk-aware and perturbation-robust decision criterion The framework predicts a compact set of futurecondition branches for each scene, performs candidate-conditioned latent consequence propagation, decodes one long-horizon outcome sample for each branch, estimates an outcome distribution through ordered quantiles, and ranks candidates using a criterion that combines the mean, lower-tail Conditional Value-at-Risk (CVaR) [36], and cross-condition instability

Distributional Outcome Estimation

The long-horizon outcome sample under condition l is defined as yˆi,l = ∑ h=1 H h−1 ri,l,h + γH rterm i,l The set yˆi,l L l=1 provides a finite-sample representation of its long-horizon outcome distribution This is achieved by pooling each branch-specific latent consequence trajectory over time and then pooling across branches Based on h˜i, the distribution head outputs M ordered quantiles, q i = h q (i)

Risk-Sensitive Ranking Criterion

To account for tail risk, RiSE-Drive uses the lower-tail Conditional Value-at-Risk (CVaR) [36] to define CVaR ↓ α (Yi) = 1/Mα ∑ m=1 Mα q(m) i The final risk-sensitive ranking score is defined as ˆJ i = μˆi − λtail Δˆ i − λvar σˆ i This formulation makes candidate selection explicitly sensitive to tail failures rather than only nominal performance

Experimental Results

RiSE-Drive achieves state-of-the-art results on NAVSIMv1 [8], NAVSIM-v2 [3], and HUGSIM [55] Under a matched DrivoR setup, it improves PDMS from 93.1 to 94.2, EPDMS from 48. p. 1

Where the paper says it

Under a matched DrivoR setup [18], it improves PDMS from 93. 2, EPDMS from 48.

3 to 51.3, and average HD-Score from 35. p. 1
Where the paper says it

2, EPDMS from 48. 3, and average HD-Score from 35.

7 to 39.5 The gain is largest on the hard and extreme subsets, where HD-Score rises from 18. p. 7
Where the paper says it

8 points, respectively. The gain is largest on the hard and extreme subsets, where HD-Score rises from 18.

8 to 22.7 and from 32.5 to 39.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Risk-sensitive candidate selection based on distributional outcomes enables preference for candidates that are not only strong under nominal conditions, but also reliable under long-horizon rollout and safety-critical deviations.

  2. The system can explicitly model and rank trajectories using statistics like the mean, lower-tail Conditional Value-at-Risk (CVaR), and cross-condition instability rather than a single deterministic score.

  3. The evaluator can identify hidden tail failures by estimating an outcome distribution through ordered quantiles, which exposes risks that are otherwise invisible to traditional point scores.

  4. The system can improve robustness on hard and perturbation-heavy settings, as shown by achieving state-of-the-art results where the gain is concentrated on Stage-II safety terms.

  5. By training with a pairwise ranking loss, the evaluator learns a criterion that maximizes the risk-sensitive score J hat i, which favors trajectories that remain reliable not only on average, but also under plausible downstream deviations and interactionheavy rollouts.

Abstract

End-to-end autonomous driving systems increasingly rely on generating multiple trajectory candidates and selecting one for execution. In this setting, a central challenge is no longer only candidate generation, but how to evaluate candidates according to their long-horizon consequences under uncertainty. Existing methods still mainly use deterministic point scores, heuristic subscores, or single-outcome evaluations, which makes it difficult to capture delayed failures, tail risk, and ranking instability under plausible future perturbations. We present RiSE-Drive, a framework for risk-sensitive long-horizon candidate evaluation in end-to-end autonomous driving. RiSE-Drive first performs candidate-conditioned latent future rollout, then estimates long-horizon outcome distributions instead of single scalar scores, and finally ranks candidates using a tail-risk-aware and perturbation-robust decision criterion. This reformulates candidate scoring from deterministic oracle-style evaluation into distributional outcome assessment with explicit risk control. As a result, the evaluator is encouraged to prefer candidates that are not only strong under nominal conditions, but also reliable under long-horizon rollout and safety-critical deviations. Experiments on the NAVSIM-v1 and NAVSIM-v2 benchmarks, as well as the closed-loop HUGSIM benchmark, show that RiSE-Drive achieves state-of-the-art performance and significantly improves the robustness of candidate selection.

Sources

Related papers