Safe Adaptive Cruise Control Under Perception Uncertainty: A Deep Ensemble and Conformal Tube Model Predictive Control Approach

Xiao Li, Anouck Girard, Ilya Kolmanovsky

Autonomous driving systems heavily depend on deep neural network-based perception to interpret their environment and support decision-making. To enhance robustness in these safety-critical settings, this paper proposes a Deep Ensemble of neural network regressors integrated with Conformal Prediction to estimate both vehicle states and associated perception uncertainties. Within the Adaptive Cruise Control framework, the method uses RGB image inputs to jointly predict states and generate non-uniform, probabilistically calibrated uncertainty bounds. A Conformal Tube Model Predictive Control scheme incorporates these uncertainty estimates to ensure probabilistic safety guarantees under exchangeability assumptions. Evaluations in a high-fidelity driving simulator demonstrate the algorithm’s effectiveness in maintaining safe following distances and accurate speed tracking, including under Out-Of-Distribution conditions.

[ArXiv]

Adaptive Cruise Control using CT-MPC + Deep Ensemble (in-distribution perception):

Using the same simulation setting as in the training phase, the ego vehicle is initialized with an unsafe distance headway, and therefore our CT-MPC applies harder braking, compared to the baseline (see video below), to enforce greater safety cautiousness. After approximately 1 second, once it re-enters the safe zone, the lead vehicle begins to decelerate. Instead of attempting to maintain the set speed, the ego vehicle adapts its speed to follow that of the lead vehicle, while consistently maintaining a safe headway distance. The baseline (see video below) demonstrates similar behavior after the initial braking phase. In our method, the safety is formally supported by the probabilistic lower bound.

Adaptive Cruise Control using the baseline method (in-distribution perception):

Adaptive Cruise Control using CT-MPC + Deep Ensemble (out-of-distribution perception):

Meanwhile, we replicate the same experiment under an out-of-distribution (OOD) scenario by setting the weather condition to “HardRainNoon”. Our Deep Ensemble detects the OOD condition and produces larger Conformal Prediction sets, shown in purple. This elevated perception uncertainty prompts our CT-MPC to apply full braking, bringing the ego vehicle to a stop and successfully avoiding a potential collision. In contrast, the baseline controller (see video below), which does not account for perception uncertainty, fails to react appropriately and collides with the lead vehicle.

Adaptive Cruise Control using the baseline method (out-of-distribution perception):