Closing the Navigation Compliance Gap in End-to-end Autonomous Driving

Hanfeng Wu; Marlon Steiner; Michael Schmidt; Alvaro Marcos-Ramiro; Christoph Stiller

cs.CVDec 11, 2025

Closing the Navigation Compliance Gap in End-to-end Autonomous Driving

Authors: Hanfeng Wu, Marlon Steiner, Michael Schmidt, Alvaro Marcos-Ramiro, Christoph Stiller

Organizations: Karlsruhe Institute of Technology Karlsruhe, Germany · BMW Group Munich, Germany

Abstract

Trajectory-scoring planners achieve high navigation compliance when following the expert's original command, yet they struggle at intersections when presented with alternative commands; over 30 percent of such commands are ignored. We attribute this navigation compliance gap to two root causes: (1) existing metrics like Ego Progress do not explicitly measure navigation adherence, diluting the gap between on-route and off-route trajectories; and (2) current datasets pair each scenario with a single command, preventing models from learning command-dependent behavior. We address the metric gap by introducing the binary Navigation Compliance metric (NAVI) and the derived Controllability Measure (CM), and the data gap with the NavControl dataset, 14,918 intersection scenarios augmented with all feasible alternative commands and routing annotations, yielding over 34,000 direction samples. Building on these, we propose NaviHydra, a trajectory-scoring planner incorporating NAVI distillation and Bird's Eye View (BEV)-based trajectory gathering for context-position-aware trajectory feature extraction. NaviHydra achieves 92.7 PDM score on NAVSIM navtest split and 77.5 CM on NavControl test split. Training with NavControl improves controllability across diverse architectures, confirming it as a broadly effective augmentation for navigation compliance. Our project page is https://hanfeng-wu.github.io/closing/.

Figures & tables

Fig. 1: Illustration of model predictions with alternative navigation command. The ego vehicle, shown in red, is at the entrance of an intersection. The original route is turning left . We demonstrate the models’ predictions under an alternative command ( turning right ). Our method responds correctly to the driving command, while Hydra-MDP and Transfuser following the original driving route, SSR deviating from the driving path.

Fig. 2: Distribution of driving commands in NavControl dataset for training/validation and testing. We present the number of both original driving commands and alternative driving commands. The augmentation provides a more balanced distribution of driving commands, where the original distribution is dominantly straight driving.

Fig. 3: Illustration of Navigation Compliance metric (NAVI): the blue lines are the lane sets on route, two evaluated trajectories colored with red and green are depicted. The green path ends on route hence has the NAVI of 1, while the red path ends off route hence it has NAVI of 0.

Fig. 4: Overview of NaviHydra framework. We encode the surrounding camera images with optionally LiDAR point clouds into a BEV feature. The BEV encoder is pretrained using a semantic segmentation task. For the trajectory decoder, we gather the corresponding grid from the BEV feature using the trajectory proposal to construct the trajectory query. The navigation command is integrated into the BEV feature to create the key and value inputs. The output of the trajectory decoder is then fed into a hydra scorer to align with each sub-score from the offline simulation.

Method	Inputs	NC $\uparrow$	DAC $\uparrow$	EP $\uparrow$	TTC $\uparrow$	C $\uparrow$	PDM Score $\uparrow$
PDM-Closed [ 29 ]	Perception GT	94.6	99.8	89.9	86.9	99.9	89.1
SSR $\dagger$ [ 16 ]	Camera	93.7	86.3	70.0	86.5	98.1	73.6
Transfuser* [ 4 ]	LiDAR & Camera	97.8	92.1	78.6	92.8	100	83.4
UniAD [ 3 ]	Camera	97.8	91.9	78.8	92.9	100	83.4
PARA-Drive [ 14 ]	Camera	97.9	92.4	79.3	93.0	99.8	84.0
Hydra-MDP- $\mathcal{V}_{4096}$ $\ddagger$ [ 1 ]	Camera	98.3	97.3	84.3	93.4	100	89.0

TABLE I: Quantitative comparison in navtest split of NAVSIM benchmark.

\dagger

: retrained on navtrain . *: evaluated using official checkpoint.

\ddagger

: retrained with our

\mathcal{V}_{4096}

trajectory proposals.

Fig. 5: Qualitative comparison of two selected scenes in navtest split. Input navigation commands are listed on the left. Routes are displayed as blue lines in the semantic map. Additionally, PDM scores along with other sub-scores for the evaluated trajectories are provided. Our trajectory achieves the highest PDM-score and demonstrates the best alignment with the navigation command.

Method	Inputs	NC $\uparrow$	DAC $\uparrow$	EP $\uparrow$	TTC $\uparrow$	C $\uparrow$	NAVI $\uparrow$	PDM Score $\uparrow$	CM $\uparrow$
SSR $\dagger$ [ 16 ]	Camera	91.3	80.6	67.7	83.3	95.4	62.8	68.8	49.9
Transfuser* [ 4 ]	LiDAR & Camera	96.7	91.7	80.6	91.6	99.8	67.5	83.4	59.4
Hydra-MDP- $\mathcal{V}_{4096}$ $\ddagger$ [ 1 ]	Camera	98.0	97.4	88.4	95.0	100	69.6	91.4	63.8
NaviHydra- $\mathcal{V}_{4096}$ (Ours)	Camera	98.4	97.8	87.1	96.0	99.2	73.9	91.4	67.6
SSR $\lx@sectionsign$ [ 16 ]	Camera	86.8	78.9	64.0	77.5	80.9	70.2	61.8	48.6
Transfuser $\lx@sectionsign$ [ 4 ]	LiDAR & Camera	88.8	88.2	73.1	81.3	89.9	81.5	72.6	62.6

TABLE II: Controllability evaluation in NavControl test split.

\dagger

: retrained on navtrain . *: evaluated using official checkpoint.

\ddagger

: retrained with our

\mathcal{V}_{4096}

trajectory proposals.

\lx@sectionsign

: trained with additional NavControl intersection scenarios.

Fig. 6: Qualitative comparison of controllability test in NavControl split. We present two scenarios, each of them is evaluated with 3 different navigation commands (Turn Left, Go Straight, Turn Right) as input. Only the scene corresponding to the original navigation command displays the ground truth trajectory, which is annotated with a green box.

ID	Modules		navtest Metrics							NavControl Metrics
ID	NAVI loss	Traj gathering	NC $\uparrow$	DAC $\uparrow$	EP $\uparrow$	TTC $\uparrow$	C $\uparrow$	NAVI $\uparrow$	PDM Score $\uparrow$	NAVI $\uparrow$	PDM Score $\uparrow$	CM $\uparrow$
1	✗	✓	98.4	98.3	88.2	95.8	100	97.8	92.1	64.8	84.5	55.8
2	✓	✗	98.4	98.2	83.9	95.5	100	98.9	89.5	85.9	84.5	73.5
3	✓	✓	98.7	98.6	88.7	96.2	100	98.0	92.7	87.6	88.0	77.5

TABLE III: Ablation studies of various model components of NaviHydra in navtest and NavControl evaluation.

Autonomous Driving

Autonomous driving develops vehicles that perceive road scenes, predict interactions, and plan and control safe movement with limited or no human input.

See all 536 papers

TrafficSignBench: Rule-Centric Closed-Loop Evaluation of Traffic-Sign Compliance in Autonomous Driving
Sep 29, 2026Victoria Smirnova, Viktoriia Zinkovich, Gregorii Bukhtuev +4Autonomous Driving Benchmarks Autonomous Driving Safety Evaluation
Lomonosov Moscow State University · FusionBrain Lab · Moscow Institute of Physics and Technology +3
SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving
Mar 31, 2026Wenchao Sun, Xuewu Lin, Keyu Chen +4Autonomous Driving Planning End-to-End Autonomous Driving
Tsinghua University, Beijing, China · Horizon Continental Technology, Beijing, China · Horizon, Beijing, China
Designing Versatile Samples for Learned Trajectory Scoring
Sep 1, 2026Yaguang Li, Jiaru Zhang, Chuheng Wei +2Autonomous Driving Datasets Autonomous Driving Planning
Purdue University · Bosch Artifical Intelligence Center

Sep 29, 2026J/K move · Enter open · S save

TrafficSignBench: Rule-Centric Closed-Loop Evaluation of Traffic-Sign Compliance in Autonomous Driving

Victoria Smirnova, Viktoriia Zinkovich, Gregorii Bukhtuev, Artem Belyaev+3

Lomonosov Moscow State University · FusionBrain Lab · Moscow Institute of Physics and Technology+3

Autonomous Driving Benchmarks Autonomous Driving Safety Evaluation Autonomous Driving Autonomous Driving Simulation Autonomous Driving Scenario Generation

Open paper AskPDF arXiv

Autonomous driving planners are typically evaluated using aggregate metrics such as driving score, destination rate, and collision rate, which do not explicitly measure compliance with traffic rules. As a result, planners can achieve high benchmark scores while still exhibiting unsafe or illegal behaviors, limiting their applicability to real-world deployment. To address this gap, we introduce TrafficSignBench, a large-scale, traffic sign-centric benchmark for systematic and interpretable evaluation of traffic-rule compliance in autonomous driving. Our framework combines real-map-based simulation for realistic road layouts with rule-targeted procedural scenario generation for scalable and balanced coverage of underrepresented rules. We implement traffic rules corresponding to 34 traffic signs, each equipped with an automatic rule checker for detecting violations during closed-loop execution. This design yields 29,000 diverse road scenes and 29 distinct testing scenario types, enabling controlled evaluation of rule-specific planner behavior. We construct 5,800 testing scenes and demonstrate that current autonomous driving planners can exhibit poor traffic-rule compliance despite strong performance on standard evaluation metrics. To address this limitation, we transform existing planners into rule-compliant trajectory experts via explicit traffic-sign constraints, enabling scalable generation of high-quality oracle trajectories for fine-tuning.

Abstract

Autonomous Driving

Explore similar work

TrafficSignBench: Rule-Centric Closed-Loop Evaluation of Traffic-Sign Compliance in Autonomous Driving

SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving

Designing Versatile Samples for Learned Trajectory Scoring