Towards Safer Autonomous Driving in an Open World: A Dual-Process Approach

Simon Janssen; Michiel Braat; Chris van der Ploeg; Serge Thill; Jan-Pieter Paardekooper

cs.AIOct 2, 2026

Towards Safer Autonomous Driving in an Open World: A Dual-Process Approach

Authors: Simon Janssen, Michiel Braat, Chris van der Ploeg, Serge Thill, Jan-Pieter Paardekooper

Organizations: Netherlands Organisation for Applied Scientific Research, Integrated Vehicle Safety Group, 5700 AT Helmond, The Netherlands · Eindhoven University of Technology, Dynamics and Control Group, Mechanical Engineering Dept., P.O. Box 513, 5600 MB, Eindhoven, The Netherlands · Radboud University, Donders Institute for Brain, Cognition and Behaviour, P.O.Box 9010, 6500 GL, Nijmegen, The Netherlands

Abstract

Before autonomous driving systems can be deployed on public roads, it is vital that these systems comply with safety standards, traffic rules, and social norms. Although neural networks trained on large amounts of driving data perform well in routine driving tasks, these models often struggle in novel situations that are not well-represented in the data. In this work, we propose a novel framework that combines a neural network for intuitive, learning-based planning in routine driving tasks with model predictive control for reasoning-based planning in unfamiliar situations, inspired by Dual Process Theory. A meta-cognitive component is designed to switch between the two, using a knowledge graph to reason about contextual risk based on explicit perceptual information and relevant traffic rules and social norms. Contextual risk is represented through risk fields, guiding both the switching mechanism in the meta-cognitive component and compliance with safety standards, traffic rules, and social norms in the reasoning-based planner. The effectiveness of our framework is tested in CARLA for variations of a typical out-of-distribution situations involving (emergency) vehicles running a red light. We show that the novel architecture reduces the number of collisions in the scenarios by 89% and improves compliance with the special right-of-way rules, compared to the NN-only planner.

Figures & tables

Fig. 1 : The overall structure of the proposed framework.

Fig. 2 : Entity hierarchy.

Fig. 3 : Most important relations for physical entities.

Fig. 4 : Most important relations for infrastructure entities.

Priority	Natural Language Rule
0	When a line is dashed, it is acceptable to cross
0	When a line is solid, it is not acceptable to cross
0	When a line is a road edge, it is not acceptable to cross
1	When a line is dashed, it is on the left side of the road and connected to a lane in the opposite direction, it is not acceptable to cross
2	When a line is a left or right border of a lane that is itself part of an intersection, it is acceptable to cross
3	When a line is a left or right border of a lane that is itself part of an intersection, and the ego vehicle is on that lane, it is not acceptable to cross

TABLE I : Natural language descriptions and priorities of rules that specify line crossing acceptability.

Priority	Natural Language Rule
0	An object has medium risk.
1	When an object is behind the ego vehicle, it has low risk
2	When an object is in front of the ego vehicle,and the object is not on a lane in the opposite direction or a lane that is in the opposite direction of another lane that the ego vehicle can reach by following its current lane, and the ego vehicle could reach the object in less than 2 seconds (assuming constant ego velocity), it has high risk.
3	When an object has an object type emergency vehicle, it has high risk.

TABLE II : Natural language descriptions and priorities of rules that specify object risk levels.

Priority	Natural Language Rule
0	It is permissible to drive on a lane.
1	When a lane is in the opposite direction of the lane that the ego vehicle is currently on, or a lane is in the opposite direction of another lane that the ego vehicle can reach by following its current lane, it is not permissible to drive on.

TABLE III : Natural language descriptions and priorities of rules that specify driving permissibility.

Priority	Natural Language Rule
0	The ego vehicle has the right of way at an intersection.
1	When an intersection is managed by traffic lights, the traffic light is turned red, and this is the first intersection that the ego vehicle approaches by following its current lane, the ego vehicle does not have the right of way.
1	When an intersection is not managed by traffic lights, it is the first intersection that the ego vehicle approaches by following its current lane, it is the first intersection that another physical entity approaches by following its current lane, the physical entity comes from the right, and either the physical entity is expected to enter the intersection within 5 seconds, or its distance to the intersection is less than 10 meters, the ego vehicle does not have right of way.
2	When an intersection is the first intersection that the ego vehicle approaches by following its current lane, and it is the first intersection that another physical entity approaches, the physical entity has object type emergency vehicle, and either the physical entity is expected to enter the intersection within 10 seconds, or the distance to the intersection is less than 20 meters, the ego vehicle does not have right of way.
3	When the ego vehicle is on a lane that is part of an intersection, the ego vehicle has right of way.

TABLE IV : Natural language descriptions and priorities of right of way rules.

Object Class	Risk Level	$A_{O}$	$\sigma_{x}$	$\sigma_{y}$
Car	Low	2	$1.5l_{O}+0.05$	$1.1w_{O}+0.05$
Car	Medium	3	$1.5l_{O}+0.1$	$1.1w_{O}+0.1$
Car	High	4	$1.5l_{O}+0.3$	$1.1w_{O}+0.3$
Pedestrian	Low	2	$1.5l_{O}+1.2$	$1.1w_{O}+1.2$
Pedestrian	Medium	4	$1.5l_{O}+1.7$	$1.1w_{O}+1.7$
Pedestrian	High	7	$1.5l_{O}+2.2$	$1.1w_{O}+2.2$

TABLE V : Risk parameters for objects, based on classification, dimensions, and risk level.

Lane Type	Acceptability	$A_{L}$	$\sigma$
Regular	Acceptable	1.5	$0.6$
Regular	Not acceptable	4	$0.6$
Intersection	Acceptable	0	$0.6$
Intersection	Not acceptable	0.65	$0.6$

TABLE VI : Risk parameters for lane markings based on the lane type and line crossing acceptability.

Precedence	$A_{I}$	$\sigma$
Right of way	0	1.5
No right of way	4	1.5

TABLE VII : Risk parameters for intersections based on right of way.

Permissibility	$A_{D}$
Permissible	0
Not permissible	3

TABLE VIII : Risk parameters for lanes based on driving permissibility.

Par.	Description	Value	Unit
$t_{s}$	Sampling time	0.15	s
$N$	Maximum trajectory horizon	20	-
$T$	Decision threshold MCC	0.7	-
$\ell_{\mathbf{u}}$	Input weights	[0.001, 1]	-
$m$	Terminal weights	[0.1, 0.01, 0, 0, 0]	-
$\underline{x},\bar{x}$	Position x-coordinate	[ $-\infty,\infty$ ]	m

TABLE IX : Simulation parameters.

Fig. 5 : For different intersection setups were used in the simulations.

Parameter	range
Scenario trigger distance	[20, 23, 27, 30] (m)
Adversary speed	[35, 40, 45, 50, 55] (km/h)
Adversary vehicle model	[ambulance, car]
Intersection	[A, C, B, D]

TABLE X : Scenario parameters.

Infraction Type	TF++	TF++ & MCC	TFv6
Collisions with Vehicle	48	5	8
Collisions with Road Layout	0	7	0
Outside lane	0	1	0

TABLE XI : Number of scenarios with infraction.

Fig. 6 : Collisions with vehicle per scenario parameter.

Fig. 7 : Snapshot where the MCC risk is above the threshold and the MPC is activated, stopping the vehicle.

Fig. 8 : Snapshot where the MCC risk is above the threshold and the MPC is activated adjusting the system I trajectory.

Fig. 9 : Snapshot where the MPC engages but only is able to calculate a path that collides with a roadside fence.

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 538 papers

READ: Learning Risk-Informed Fields for End-to-End Autonomous Driving
Sep 11, 2026Zhiyuan Liu, Yuanxin Tian, Zehong Ke +5End-to-End Autonomous Driving Safe Motion Planning
School of Vehicle and Mobility, Tsinghua University.
Data-Driven Risk Fields for Safer End-to-End Autonomous Driving
Sep 9, 2026Yuanxin Tian, Zhiyuan Liu, Jinhao Li +9End-to-End Autonomous Driving Safe Motion Planning
Tsinghua University · Yinwang · KAIST
Reason--Imagine--Act: Closed-Loop LLM Decision Making with World Models for Autonomous Driving
May 19, 2026Zhengqi Sun, Yiwen Sun, Boxuan Liu +3Language Model-Based Control World Models
Department of Information Management, Peking University, Beijing 100871, China · School of Intelligence Science and Technology, Peking University, Beijing 100871, China · State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing 100080, China +3

Sep 11, 2026J/K move · Enter open · S save

READ: Learning Risk-Informed Fields for End-to-End Autonomous Driving

Zhiyuan Liu, Yuanxin Tian, Zehong Ke, Jinhao Li+4

School of Vehicle and Mobility, Tsinghua University.

End-to-End Autonomous Driving Safe Motion Planning

Open paper AskPDF arXiv

Autonomous driving requires more than recognizing what is present in a scene: a planner must determine how road structure, surrounding agents, and their motion states should influence a future maneuver. Existing learning-based planners can capture these influences through latent scene features and trajectory decoders, but the relationship between environmental factors and candidate actions often remains implicit. This limits the ability to inspect, diagnose, or refine how scene context affects the safety of a predicted trajectory. Classical safety fields provide an explicit spatial representation of this relationship, but their risk shapes and relative weights are prescribed in advance and do not adapt to each scene. We introduce READ, a framework that learns an explicit, planning-aligned risk representation from complementary geometric and behavioral constraints. READ instantiates this representation as a continuous spatiotemporal field, enabling differentiable queries along candidate trajectories. The learned field connects scene understanding with action selection by encouraging predicted trajectories to align with low-risk regions, while retaining a differentiable interface for trajectory evaluation and refinement. READ integrates with both end-to-end planners and Vision-Language-Action models. Experiments on NAVSIM show consistent gains across matched end-to-end backbones and strong performance in a VLA setting; READ also achieves competitive results on NAVSIM v2. These results establish learned spatial risk as an explicit, adaptable representation for safe planning.

Abstract

Autonomous Driving

Explore similar work

READ: Learning Risk-Informed Fields for End-to-End Autonomous Driving

Data-Driven Risk Fields for Safer End-to-End Autonomous Driving

Reason--Imagine--Act: Closed-Loop LLM Decision Making with World Models for Autonomous Driving