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.
Figures & tables
Figure 1. Why deterministic mean scoring can fail. Two ego candidates may have similar mean scores, yet one can still hide a small set of catastrophic future-condition branches. RiSE-Drive therefore evaluates each candidate under multiple future-condition branches, forms a distribution over branch-level long-horizon outcomes, and ranks candidates with risk-sensitive statistics such as the mean, lower quantiles, and CVaR. RiSE-Drive evaluates each candidate under multiple future-condition branches, rolls each branch forward to obtain one long-horizon outcome sample, forms a distribution over branch outcomes, and ranks candidates using risk-sensitive statistics including the mean, lower quantiles, and CVaR.
Figure 2. Overview of the RiSE-Drive pipeline. Multi-view observations are encoded, and the proposal generator produces K ego-trajectory candidates. A future-condition generator predicts L branch-specific condition codes that parameterize plausible downstream evolutions. For each candidate–condition pair, RiSE-Drive initializes a latent consequence state, propagates it along the trajectory, and decodes one branch-level long-horizon outcome sample. The resulting branchwise consequence traces are pooled over time and across branches and mapped to ordered outcome quantiles by a distribution head. The final ranking score combines the mean estimate, lower-tail CVaR, and cross-condition instability, thereby favoring candidates that remain reliable under plausible future deviations rather than merely optimal under a single nominal forecast. Overview of RiSE-Drive. Multi-view observations are encoded into scene features. The proposal generator creates candidate ego trajectories, while the future-condition generator creates plausible downstream branches. Each candidate and branch pair is propagated to obtain a long-horizon outcome sample. The resulting distribution is summarized with ordered quantiles and ranked using the mean, lower-tail CVaR, and cross-condition instability.
Method
Input
Backbone
Venue
NC ↑
DAC ↑
EP ↑
TTC ↑
Conf. ↑
PDMS ↑
Human
-
-
-
100
100
87.5
100
99.9
94.8
DrivingGPT ( Chen et al., 2025 )
C
VQ-VAE/GPT
ICCV 2025
98.9
90.7
79.7
94.9
95.6
82.4
VADv2 ( Chen et al., 2024 )
C
ResNet-34
arXiv 2024
97.9
91.7
77.6
92.9
100
83.0
UniAD ( Hu et al., 2023 )
C
ResNet-34
CVPR 2023
97.8
91.9
78.8
92.9
100
83.4
LTF ( Chitta et al., 2023 )
C
ResNet-34
TPAMI 2022
97.4
92.8
79.0
92.4
100
83.8
TransFuser ( Chitta et al., 2023 )
C & L
ResNet-34
TPAMI 2022
97.7
92.8
79.2
92.8
100
84.0
Table 1. NAVSIM-v1 evaluation results on the Navtest split. C denotes camera-only input and L denotes LiDAR.
Stage 1
Stage 2
Method
Backbone
NC
DAC
DDC
TLC
EP
TTC
LK
HC
EC
NC
DAC
DDC
TLC
EP
TTC
LK
HC
EC
EPDMS
LTF ( Chitta et al., 2023 )
ResNet-34
96.2
79.5
99.1
99.5
84.1
95.1
94.2
97.5
79.1
77.7
70.2
84.2
98.0
85.1
75.6
45.4
95.7
75.9
23.1
DiffusionDrive ∗ ( Liao et al., 2025 )
ResNet-34
96.0
79.7
97.4
99.5
81.3
93.1
90.8
96.8
73.8
82.1
72.2
88.5
98.7
85.1
78.8
49.2
89.3
71.2
24.2
GuideFlow ( Liu et al., 2025 )
ResNet-34
96.6
80.5
96.3
99.3
82.3
94.9
91.5
97.7
67.8
87.3
76.7
88.8
99.2
84.3
85.1
49.7
93.1
44.5
27.1
RAP-DINO ( Feng et al., 2026 )
ViT-H
97.1
94.4
98.8
99.8
83.9
96.9
94.7
96.4
66.2
83.2
83.9
87.4
98.0
86.9
80.4
52.3
95.2
52.4
39.6
DriveSuprim ( Yao et al., 2025 )
V2-99
98.9
95.1
99.2
99.6
76.1
99.1
94.7
97.6
54.2
87.9
88.8
89.6
98.8
80.3
86.0
53.5
97.1
56.1
42.1
Table 2. NAVSIM-v2 navhard-two-stage results. ∗ denotes results reproduced with the official code repository or official checkpoint.
Method
RC
HD-Score
E
M
H
X
Avg.
E
M
H
X
Avg.
VAD ( Jiang et al., 2023 )
51.3
31.1
25.3
26.5
31.4
36.3
9.5
8.0
11.5
13.4
LTF ( Chitta et al., 2023 )
67.8
35.1
26.2
40.5
38.9
58.9
18.0
9.8
25.9
23.7
UniAD ( Hu et al., 2023 )
78.4
60.5
33.6
17.8
45.9
64.9
45.8
20.6
6.6
32.7
DrivoR ( Kirby et al., 2026 )
80.9
50.5
33.8
47.1
49.8
73.3
34.6
18.8
32.5
35.7
Ours
76.5
53.3
34.1
46.3
50.1
74.6
37.2
22.7
39.2
39.5
Table 3. Photorealistic closed-loop evaluation on HUGSIM.
Figure 3. Qualitative results of RiSE-Drive on representative NavSafe safety-critical scenarios. From left to right, columns (a)–(d) correspond to UNPL ( unprotected left turn ), EXR ( exit ramp ), YLD ( yielding ), and RDBT ( roundabout ). Each column shows two temporal snapshots of one representative case. Magenta denotes the RiSE-Drive plan, green the human future trajectory, orange dashed the ego history, and blue the future trajectories of surrounding vehicles. Across all four cases, RiSE-Drive produces smooth, route-consistent, and interaction-aware behaviors while remaining close to safe human driving in conflict-prone regions. Four columns show unprotected-left-turn, exit-ramp, yielding, and roundabout scenarios, with two temporal snapshots in each column. Magenta curves are RiSE-Drive plans, green curves are human future trajectories, orange dashed curves are ego histories, and blue curves are future trajectories of surrounding vehicles.
NAVSIM-v1
NAVSIM-v2 NavHard
Configuration
PDMS ↑
NC ↑
EPDMS ↑
TTC ↑
B-I: DrivoR
93.1
98.9
48.3
88.0
B-II: DrivoR w/o SD
91.9
98.4
45.6
84.6
+ FCG
92.6
98.7
46.8
86.5
+ CTE
92.9
98.8
47.5
87.3
+ SLE + LCP
93.5
99.1
49.2
89.6
Table 4. Component-wise ablation of the proposed evaluator. Starting from the reduced baseline B-II, each row adds one RiSE-Drive component to the configuration immediately above.
Configuration
Optim Params (M)
Total Params (M)
PDMS ↑
EPDMS ↑
HD-Score Avg. ↑
DrivoR
18.8
41.8
93.1
48.3
35.7
B-II
17.2
39.9
91.9
45.6
33.8
RiSE-Drive
18.9
42.3
94.2
51.3
39.5
Table 5. Parameter-efficiency of RiSE-Drive relative to DrivoR and the reduced baseline.
Safety is a fundamental requirement for autonomous driving, yet existing end-to-end driving models still lack explicit risk-aware learning capacities. Existing rule-based risk models provide interpretable safety priors, yet their absolute risk scores depend on handcrafted functions, coefficients, and thresholds. Learning-based risk representations reduce part of this manual design, but their supervision often relies on occupancy-derived labels or heuristic cost values, which may not capture ego-conditioned planning risk. In this paper, we propose DRiF, a data-driven risk-field framework for safer end-to-end autonomous driving. DRiF learns a shared BEV feature with static map segmentation, dynamic risk prediction, and vehicle planning. For dynamic risk learning, DRiF converts rule-based safety priors into pairwise risk labels, and trains the risk field to preserve relative risk ordering instead of regressing handcrafted absolute scores. Experiments on Bench2Drive show that DRiF achieves competitive overall performance, with consistent improvements in driving score, success rate, and collision-related metrics. These results establish relative risk supervision as an effective way to connect explicit safety structure with end-to-end planning. The data and code will be publicly available.
Candidate-based end-to-end driving planners reason over a finite set of trajectory hypotheses. However, for sparse planners, a limited candidate set may provide insufficient support for scene-relevant maneuvers, while fixed downstream candidate budgets do not account for prediction uncertainty. To address these challenges, we introduce Expand-and-Prune, a test-time framework that jointly adapts candidate support and capacity without updating learned parameters. The framework expands the native candidate set using semantic driving intent and temporally propagated predictions and then uses uncertainty in intermediate trajectory scores to adaptively prune the resulting candidates. Across four heterogeneous candidate-based planners and four benchmarks, Expand-and-Prune consistently improves planning performance. It achieves state-of-the-art results on LangAuto, increasing the Driving Score from 68.2 to 75.9, while providing further improvement on Bench2Drive, Fail2Drive, and NAVSIMv2, with particularly strong improvements in out-of-distribution scenarios. Moreover, in its full instantiation, 72.5% of selected trajectories originate from the newly introduced intent-guided or temporal candidates, confirming that candidate-space expansion provides useful hypotheses beyond those generated natively by the planner. The code is available at https://github.com/iis-esslingen/Expand-and-Prune.
Fabian Schmidt, Karol Fedurko, Markus Enzweiler +1
Esslingen University of Applied Sciences · University of Freiburg
Multimodal trajectory prediction improves behavioral coverage in end-to-end autonomous driving, but existing methods remain limited by sparse scene representations. Incomplete evidence leads to low-quality candidate generation and unreliable ranking among geometrically similar trajectories. On a register-based baseline, bad and poor candidates constitute 19.74% of the candidate set, while the oracle-best candidate ranks only 33.9th on average. We propose RRDrive, which introduces risk-aware occupancy as a dense, temporally aligned, and trajectory-queryable representation. Its global structure guides high-quality multimodal generation, while candidate-conditioned risk queries support fine-grained selection. We further construct RiskOcc4D-NAVSIM with automatic risk annotations. RRDrive achieves a selected-trajectory PDMS of 0.951, representing a 1.5% relative improvement over the baseline (0.937), and improves the average candidate PDMS by 7.7%. In challenging scenes, it improves candidate PDMS by 30.2% and increases the Spearman correlation among good candidates by 0.41, from 0.26 to 0.67. To move beyond this oracle setting, we further develop an external RiskOcc predictor, a perception module that estimates risk-aware occupancy directly from sensor inputs. The competitive performance validates the representation's feasibility.
Jiaxing Chen, Hengduo Zou, YuKai Qin +3
School of Vehicle and Mobility, Tsinghua University · The Hong Kong University of Science and Technology (Guangzhou) · Neolix, Beijing, China