PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.04088v1cs.AIcs.RO

TL;DR

Neural-network planners can struggle in novel driving situations, where semantically irrelevant patterns may lead to overconfident or incorrect predictions. The paper combines neural planning with context-aware MPC, using a meta-cognitive component and risk fields to assess trajectories and switch planners. In CARLA red-light scenarios, the framework reduces vehicle-to-vehicle collisions by 89% compared with the NN-only planner.

  • Problem

    Neural networks often struggle to generalize safely to novel inputs, which may yield overconfident or incorrect predictions based on semantically irrelevant patterns.

    Source: §I · p. 1

  • Method

    A meta-cognitive component uses knowledge-graph-informed risk fields to assess neural-planner trajectories and switch to MPC, which uses those fields for reasoning-based planning.

    Source: §I · p. 1, 4) Evaluating Predicted Path · p. 5, §V · p. 7

  • Results

    89% fewer vehicle-to-vehicle collisions in CARLA simulations involving an adversary vehicle running a red light at intersections.

    Source: §V · p. 7

  • Takeaways & Limitations

    In the tested simulations, risk-triggered MPC switching improved safety and adaptivity while offering a safer, less human-like alternative trajectory.

    Source: §V · p. 7

    Evaluation covers CARLA simulations of adversary vehicles running red lights at intersections; future work includes scaling to other traffic scenarios.

    Source: §V · p. 7

Abstract

from arXiv · show

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.

I. INTRODUCTION

Neural-network planners can struggle to generalize safely to novel situations, where semantically irrelevant patterns may produce overconfident or incorrect predictions. The proposed framework combines neural planning with reasoning-based MPC and a meta-cognitive component that switches between them using risk fields.

I. INTRODUCTION

  • Novel inputs can prompt neural networks to rely on semantically irrelevant patterns, producing overconfident or incorrect predictions in situations where safe behavior is essential.Source: §I · p. 1
  • The framework uses an end-to-end network for familiar situations and a context-aware MPC planner for out-of-distribution situations, guided by principles from Dual Process Theory.Source: §I · p. 1
  • A meta-cognitive component evaluates the neural planner’s trajectory for safety and social desirability using risk fields, then decides whether to switch to the reasoning-based planner.Source: §I · p. 1
  • The framework’s effectiveness is demonstrated through simulation experiments in CARLA.Source: §I · p. 1

II. RELATED WORK

Prior work explores end-to-end driving and safety layers, while trajectory evaluation offers a way to judge individual predictions by risk or reward. This work uses knowledge-graph-shaped risk fields to score contextual safety and desirability for switching planners.

A. End-to-End Autonomous Driving

B. Safety in Autonomous Driving

  • Safety approaches integrate explicit rules or cost functions into planners or use external layers, while this work uses MPC and a knowledge graph rather than VLMs for context-aware trajectories.Source: B. Safety in Autonomous Driving · p. 2

C. Trajectory Evaluation

  • Trajectory scoring can assess actual risk or reward rather than approximate novelty or uncertainty, and this work shapes risk fields with knowledge-graph context to capture traffic rules and social norms.Source: C. Trajectory Evaluation · p. 2
  • The proposed scoring function excludes comfort and route progression because they are not treated as relevant reasons to switch systems in this work.Source: C. Trajectory Evaluation · p. 2

III. METHOD

The framework passes a neural planner’s initial solution to a meta-cognitive component, which evaluates it and may hand control to MPC. The MCC uses knowledge-graph context to assess safety, legality, and social desirability; its perceptual inputs combine model predictions with simulator ground truth.

III. METHOD

  • The neural planner predicts an initial solution from camera, LiDAR, and GPS inputs; when engaged, MPC uses risk fields to produce safer, more socially desirable trajectories at the expense of human-like behavior.Source: §III · p. 2

A. System I (Neural Network)

  • TransFuser++ integrates front-facing RGB camera and 360° LiDAR inputs, then predicts a target speed and ten future path positions conditioned on a GPS goal and current speed.Source: A. System I (Neural Network) · p. 2

B. Meta-Cognitive Component

  • The MCC evaluates the neural planner’s path for safety, legality, and social desirability using contextual information derived from a knowledge graph.Source: B. Meta-Cognitive Component · p. 2

1) Input Description:

  • A separate perception module is outside the study’s scope; MCC inputs instead combine TransFuser++ stochastic bounding-box predictions with supplementary ground-truth data from perfect simulation sensors.Source: 1) Input Description · p. 3

2) Knowledge Graph:

The knowledge graph combines perceptual information with inferred traffic rules and social norms to represent driving context. Prioritized default rules address lane-crossing acceptability, object risk, lane permissibility, and intersection precedence.

2) Knowledge Graph:

  • Perceptual information and inferred traffic rules and social norms are linked in a knowledge graph to derive implicit contextual information.Inference rules determine such things as lane-marking crossing acceptability, object risk, lane permissibility, and intersection precedence.Source: 2) Knowledge Graph · p. 3
  • The ontology represents concepts as entities, their interactions as relations with roles, and properties such as speed or position as attributes.Examples of entities include detected objects, lanes, and traffic signs; an “in front” relation links an ego subject and an actor object.Source: 2) Knowledge Graph · p. 3
  • Rules for line-crossing acceptability, object risk, lane permissibility, and intersection precedence are implemented as prioritized defaults to handle conflicts, exceptions, and incomplete knowledge.At inference time, the highest-priority rule is selected; the latter two rule groups are introduced for this use case.Source: 2) Knowledge Graph · p. 4

3) Populating Risk Fields:

Risk fields represent both object-imposed and infrastructure risk. Object risk is modeled with Gaussian distributions, while infrastructure risk maps category masks directly or smooths selected categories with Gaussian convolution.

3) Populating Risk Fields:

  • Object-imposed risk is represented by a bivariate Gaussian distribution, with amplitude and standard deviation set by object classification, dimensions, and risk level.The parameter table is organized by object classification, dimensions, and risk level.Source: 3) Populating Risk Fields · p. 4, Table V · p. 4
  • Infrastructure risk maps lane and off-road masks directly, while lane-boundary and intersection masks are convolved with a 2D Gaussian to form smooth risk maps.Each infrastructure category is encoded as a binary mask over the environmental map, relative to the ego vehicle.Source: 3) Populating Risk Fields · pp. 4–5
  • The infrastructure fields represent risk from crossing lane boundaries, driving at intersections, driving in lanes, and driving off-road.Risk parameters are organized for lane markings, intersections, and lanes by crossing acceptability, right-of-way, and driving permissibility, respectively.Source: 3) Populating Risk Fields · p. 5, Table VI · p. 5, Table VII · p. 5

4) Evaluating Predicted Path:

The predicted neural-network path is converted into a time-based ego trajectory, evaluated against contextual risk fields, and used to trigger a switch to MPC when peak risk exceeds a threshold.

4) Evaluating Predicted Path:

  • The neural-network path is converted into a time-based trajectory using cubic-spline interpolation, simulated ego motion, and a speed-dependent horizon limited by the predicted 12.5-meter path.A longitudinal PID controller predicts throttle and braking, a learned polynomial model estimates future velocities, and surrounding trajectories use constant-velocity or constant-steering-angle-and-velocity models.Source: 4) Evaluating Predicted Path · p. 5
  • Risk along the predicted trajectory is computed by adding the relevant risk-field values at the ego vehicle’s positions over time.The trajectory records ego position at each time step across the prediction horizon.Source: 4) Evaluating Predicted Path · p. 5
  • The MCC switches to the MPC-based planner when the maximum risk along the trajectory exceeds an empirically determined threshold.The maximum is taken across the trajectory’s time steps.Source: 4) Evaluating Predicted Path · p. 5

C. System II (Model Predictive Control)

The MPC planner optimizes a minimal-risk path toward the goal using a nonlinear kinematic bicycle model and the contextual risk fields. Those fields encode the traffic rules and social norms incorporated into its objective.

C. System II (Model Predictive Control)

  • The MPC planner optimizes a path of minimal risk toward the goal point using risk fields as the basis for its objective.The global planner supplies the planar position of the reference state.Source: C. System II (Model Predictive Control) · p. 5
  • Its objective combines a nonlinear kinematic bicycle model, the current state, risk fields, and a reference state, with input bounds based on ego-vehicle properties.The model is described as suitable for urban driving.Source: C. System II (Model Predictive Control) · p. 5
  • The planner uses CasADi with the IPOPT nonlinear solver and two PID controllers for low-level vehicle control.It optimizes trajectories over the maximum prediction horizon.Source: C. System II (Model Predictive Control) · p. 5

A. Experimental Setup

The CARLA simulations test red-light-running adversaries across varied intersection conditions, with the framework implemented in Python.

A. Experimental Setup

  • Risk fields used a 1024x1024 grid for a 64m x 64m area, downscaled and smoothed for MPC; execution demands made simulation slower than real time.The grid was downscaled by a factor of eight and smoothed with cubic spline interpolation.Source: A. Experimental Setup · p. 6
  • 160 test scenarios covered normal and emergency vehicles running red lights, varying ego distance, adversary speed, and intersection, with TransFuser v6 as a baseline.The scenarios were designed so the ego vehicle was likely to collide without preventive action.Source: A. Experimental Setup · p. 6
  • Fig. 5 shows the intersection setups, while Tables IX, X, and XI list simulation parameters, scenario parameters, and scenarios with infractions.Source: Figure 5 · p. 6, Table IX · p. 6, Table X · p. 6

B. Results

The MCC substantially reduced vehicle-to-vehicle collisions, while results also reveal a steering-feasibility limitation in some interventions.

B. Results

  • 48 of 160 simulations collided without MCC, versus 5 with MCC; the TFv6 baseline had 8 collisions.The collision counts are reported in Table XI; Fig. 6 breaks vehicle-to-vehicle collisions down by scenario parameter.Source: B. Results · p. 6, Figure 6 · p. 7
  • When an ambulance approaches or an adversary crosses, MCC risk can exceed its threshold, prompting MPC to take control and stop the vehicle.Fig. 7 illustrates an intervention that stops the vehicle.Source: B. Results · p. 6, Figure 7 · p. 7
  • At intersection D, where the adversary approaches from the front, both TransFuser++ variants perform worst; all seven MCC road-layout collisions occur in Scenario 4.The MPC trajectory can demand a rotation the vehicle cannot follow, causing lane departure and collision with curbside trees or fences.Source: B. Results · pp. 6–7

V. CONCLUSIONS

The conclusion presents a dual-process planner that switches from a neural planner to context-aware MPC when trajectory risk exceeds a threshold. In CARLA red-light scenarios, it reduced vehicle-to-vehicle collisions by 89%.

V. CONCLUSIONS

  • The MCC evaluates neural-network trajectories using risk fields shaped by perceptual information, traffic rules, social norms, and predicted trajectories of surrounding objects.Source: §V · p. 7
  • Fig. 9 depicts an MPC-generated path that collides with a roadside fence, illustrating a road-layout collision despite intervention.Source: Figure 9 · p. 7
  • 89% fewer vehicle-to-vehicle collisions were reported in CARLA simulations of an adversary running a red light at intersections.Future work includes more advanced motion prediction, a more principled MCC threshold, faster computation, and scaling to other traffic scenarios.Source: §V · p. 7

Related papers

Loading 2610.04088v1…