Dual-process framework cuts red-light-running collisions from 48 to 5 in autonomous driving tests

A hybrid architecture is proposed that combines an end-to-end neural network planner with a symbolic-reasoning-based model predictive control planner, the former producing human-like trajectories in familiar situations and the latter explicitly accounting for safety standards, traffic rules, and social norms in out-of-distribution situations. Relative to an NN-only planner, the architecture adds an explicit reasoning-based safety layer aimed at novel situations that are not well-represented in the training data. Implemented and tested in CARLA, covering one out-of-distribution scenario category of red-light running with both a normal car and an emergency vehicle as the adversary.

A meta-cognitive component evaluates the safety and social desirability of the neural network's predicted trajectory using contextual risk fields derived from knowledge-graph reasoning, and triggers a switch when the maximum risk along the trajectory exceeds an empirically determined threshold of 0.7. Unlike prior switching approaches based on approximate measures such as novelty or uncertainty, this component scores trajectories on an actual risk measure and can detect more complex traffic rules and social norms. The knowledge graph is implemented in TypeDB 3.0, with rules encoded as prioritized default rules covering lane-marking crossing acceptability, object risk levels, lane permissibility, and intersection precedence.

Across 160 parameter combinations of red-light-running scenarios in CARLA, the system with the meta-cognitive component produced only 5 vehicle-to-vehicle collisions, versus 48 for TransFuser++ without it and 8 for the TFv6 comparison baseline. Relative to the NN-only planner, vehicle-to-vehicle collisions fell by 89%, and compliance with special right-of-way rules improved. Testing systematically varied trigger distance, adversary speed, adversary vehicle model, and intersection, and included a state-of-the-art baseline for comparison.

The system with the meta-cognitive component recorded 7 collisions with road layout and 1 instance of driving outside the lane, all occurring when the meta-cognitive component intervened by steering away from the adversary, because the MPC-generated trajectory required a larger rotation angle than the ego vehicle could feasibly follow at that moment. This surfaces a matching issue between the planner's intended steering angle and the angle actually executed by the PID controller during reasoning-based interventions. The phenomenon is localized to Scenario 4 in the simulations, with an example timestamp of the intervention.

The framework targets the setting of autonomous driving systems handling out-of-distribution situations at urban intersections, particularly an adversary vehicle running a red light, including an ambulance. It enables researchers and engineering teams to encode explicit traffic rules and social norms into risk fields and to switch between a neural network planner and a model predictive control planner on that basis. The meta-cognitive component and the MPC planner share the same risk fields, so updating the rules changes both the switching criterion and the safety planning together. The results apply to validation in simulation, and the authors note future work will integrate more advanced motion prediction models, adopt a more principled approach to deriving the meta-cognitive decision threshold, improve computational time, and scale the method to other traffic scenarios.

The meta-cognitive decision threshold of 0.7 is empirically determined, and the authors list a more principled derivation of this threshold as future work, so its applicability in other scenarios remains an open question. Perception input mixes stochastic bounding-box predictions from TransFuser++ with ground-truth data from perfect sensors in simulation, and the distinction between regular and emergency vehicles is based on ground truth, leaving the influence of perception uncertainty on the switching criterion to be clarified. Because of the computational demands of the meta-cognitive component and the MPC optimization, the effective simulation timestep is dictated by framework execution time, running slower than real time, so real-time feasibility remains open. The 7 road-layout collisions and 1 instance of driving outside the lane in the system with the meta-cognitive component point to a matching issue between the planner's intended steering angle and the angle actually executed by the PID controller, which the authors also describe as showing the challenge of correctly tuning the MPC. In addition, experiments cover only the red-light-running out-of-distribution scenario, so performance in other out-of-distribution situations remains to be examined.