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

Hanfeng Wu^1,2, Marlon Steiner¹, Michael Schmidt², Alvaro Marcos-Ramiro^2,*, Christoph Stiller¹

¹Karlsruhe Institute of Technology, ²BMW Group
^* denotes corresponding author

IROS 2026

NaviHydra responds correctly to an alternative navigation command at an intersection

The ego vehicle (red) is at the entrance of an intersection. The original route is turning left; we issue an alternative command (turning right). NaviHydra responds correctly, while Hydra-MDP and Transfuser keep following the original route and SSR deviates from the driving path.

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% 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 the NAVSIM navtest split and 77.5 CM on the NavControl test split. Training with NavControl improves controllability across diverse architectures, confirming it as a broadly effective augmentation for navigation compliance.

Contributions

We identify the navigation compliance gap in trajectory-scoring planners: the closest existing metric, Ego Progress, does not explicitly measure navigation adherence. Combined with the lack of intersection-level diversity in existing datasets, this gap prevents models from learning or being evaluated on command-following behavior.
We construct the NavControl dataset, which augments 14,918 intersection scenarios with permissible alternative navigation commands and routing annotations, serving as both a training augmentation and a test benchmark for navigation-aware planning.
We introduce the NAVI metric, a binary endpoint-based check for route adherence that provides the clear supervision signal that Ego Progress lacks, and define the Controllability Measure (CM) as NAVI × PDM, rewarding only trajectories that both follow the commanded direction and drive safely.
We propose trajectory gathering for context-position-aware trajectory feature extraction, which together with NAVI extends Hydra-MDP to the NaviHydra framework that achieves state-of-the-art performance on both the NAVSIM and the NavControl benchmark.

Navigation Compliance metric (NAVI)

The blue lines are the lane sets on route; two evaluated trajectories are colored red and green. The green path ends on route, hence has NAVI of 1, while the red path ends off route, hence has NAVI of 0. NAVI provides a binary, absolute route-adherence signal that Ego Progress lacks.

NaviHydra framework

Overview of NaviHydra. We encode the surrounding camera images, optionally with 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, including the new NAVI distillation head.

NAVSIM navtest results

Quantitative comparison on the navtest split of the NAVSIM benchmark. †: retrained on navtrain. *: evaluated using official checkpoint. ‡: retrained with our V₄₀₉₆ trajectory proposals.

Method	Inputs	NC↑	DAC↑	EP↑	TTC↑	C↑	PDM Score↑
PDM-Closed	Perception GT	94.6	99.8	89.9	86.9	99.9	89.1
SSR†	Camera	93.7	86.3	70.0	86.5	98.1	73.6
Transfuser*	LiDAR & Camera	97.8	92.1	78.6	92.8	100	83.4
UniAD	Camera	97.8	91.9	78.8	92.9	100	83.4
PARA-Drive	Camera	97.9	92.4	79.3	93.0	99.8	84.0
Hydra-MDP-V₄₀₉₆‡	Camera	98.3	97.3	84.3	93.4	100	89.0
Hydra-MDP++ V₈₁₉₂	Camera	98.6	98.6	85.7	95.1	100	91.0
Hydra-MDP-V₁₆₃₈₄*	Camera	98.5	98.8	85.4	94.9	100	90.8
NaviHydra-V₄₀₉₆ (Ours)	Camera	98.4	97.2	84.0	95.1	99.5	89.3
NaviHydra-V₄₀₉₆ (Ours)	LiDAR & Camera	98.7	98.6	88.7	96.2	100	92.7

NAVSIM qualitative comparison

Qualitative comparison of two selected scenes in the navtest split. Input navigation commands are listed on the left. Routes are displayed as blue lines in the semantic map. 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.

NavControl controllability results

Controllability evaluation on the NavControl test split. †: retrained on navtrain. *: official checkpoint. ‡: retrained with our V₄₀₉₆ trajectory proposals. §: trained with additional NavControl intersection scenarios. CM = NAVI × PDM.

Method	Inputs	NC↑	DAC↑	EP↑	TTC↑	C↑	NAVI↑	PDM↑	CM↑
SSR†	Camera	91.3	80.6	67.7	83.3	95.4	62.8	68.8	49.9
Transfuser*	LiDAR & Camera	96.7	91.7	80.6	91.6	99.8	67.5	83.4	59.4
Hydra-MDP-V₄₀₉₆‡	Camera	98.0	97.4	88.4	95.0	100	69.6	91.4	63.8
NaviHydra-V₄₀₉₆ (Ours)	Camera	98.4	97.8	87.1	96.0	99.2	73.9	91.4	67.6
SSR§	Camera	86.8	78.9	64.0	77.5	80.9	70.2	61.8	48.6
Transfuser§	LiDAR & Camera	88.8	88.2	73.1	81.3	89.9	81.5	72.6	62.6
Hydra-MDP-V₄₀₉₆‡§	Camera	95.8	97.2	87.7	90.3	98.2	74.8	88.3	66.5
NaviHydra-V₄₀₉₆§ (Ours)	Camera	97.2	97.4	83.2	94.3	96.4	87.6	88.0	77.5

Augmenting training with NavControl intersection scenarios (§) consistently improves NAVI across all methods. NaviHydra§ benefits the most (+9.9 CM), demonstrating that our framework is specifically designed to exploit the contrastive intersection-level signal.

NavControl qualitative comparison

Qualitative comparison on the NavControl split

Qualitative comparison of the controllability test in the NavControl split. We present two scenarios, each 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, annotated with a green box. NaviHydra reacts correctly to every command, while Hydra-MDP fails to respond and Transfuser/SSR fall short of drivable-area compliance.

BibTeX

@inproceedings{wu2026closing,
  title     = {Closing the Navigation Compliance Gap in End-to-end Autonomous Driving},
  author    = {Wu, Hanfeng and Steiner, Marlon and Schmidt, Michael and Marcos-Ramiro, Alvaro and Stiller, Christoph},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026}
}