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

TL;DR

This work tackles perception uncertainty in RGB-based adaptive cruise control by integrating a Deep Ensemble of DNN regressors with Conformal Prediction to obtain statistically guaranteed uncertainty estimates from image data. The estimated uncertainty informs a Conformal Tube Model Predictive Controller, which optimizes acceleration over a horizon while ensuring probabilistic safety bounds. Key contributions include memory-efficient ensemble training with pruning, robust uncertainty quantification under adversarial and out-of-distribution conditions, and a Conformal Tube MPC that yields a safety guarantee of at least $(1-2\alpha)$ over the prediction horizon. The approach is validated in a high-fidelity Carla simulator, demonstrating effective speed tracking, safe headway maintenance, and resilience to perception disturbances with formal probabilistic safety guarantees.

Abstract

Autonomous driving heavily relies on perception systems to interpret the environment for decision-making. To enhance robustness in these safety critical applications, this paper considers a Deep Ensemble of Deep Neural Network regressors integrated with Conformal Prediction to predict and quantify uncertainties. In the Adaptive Cruise Control setting, the proposed method performs state and uncertainty estimation from RGB images, informing the downstream controller of the DNN perception uncertainties. An adaptive cruise controller using Conformal Tube Model Predictive Control is designed to ensure probabilistic safety. Evaluations with a high-fidelity simulator demonstrate the algorithm's effectiveness in speed tracking and safe distance maintaining, including in Out-Of-Distribution scenarios.

Figures (7)

• Figure 1: A schematic diagram of the Adaptive Cruise Control implementation.
• Figure 2: Adaptive cruise controller pipeline: The Deep Ensemble uses a diverse set of DNNs for estimation, incorporating uncertainty quantification via Conformal Prediction. The Perception Uncertainty-Aware Conformal Tube MPC then computes the acceleration commands.
• Figure 3: Memory and MAE statistics of three DNN paths during the iterative pruning process: (a) Number of parameters in each DNN path. (b) Memory required for parameter storage in megabytes (MB). (c) MAE after initial training and after each iteration of fine-tuning post-pruning.
• Figure 4: Distance headway estimation results with the ground truth $d_k$ trajectory shown in red: Mean $\mu_i$ and variance $\sigma_i^2$ estimations from the three DNNs $(i=1, 2, 3)$ are visualized with colored solid lines and bands. Panels (a) and (c) display results for pre-pruning and post-pruning DNNs, respectively. The ensemble results are visualized in panels (b) and (d), with the Conformal Prediction sets $\mathcal{C}(I_{l,k}, I_{r,k})$ of different coverage rates $(1-\alpha)$ plotted as purple bands along the trajectory $\mu_k$.
• Figure 5: Uncertainty quantification using Deep Ensemble against FGSM adversarial attacks: (a) and (b) show examples of FGSM attackers on the left and right images, respectively. (c), (d), and (e) depict Conformal Prediction sets $\mathcal{C}(I_{l,k}, I_{r,k})$ with different coverage rates $(1-\alpha)$ under FGSM attacks of varying magnitudes $\epsilon$. (f) shows the size of the Conformal Prediction sets required to achieve a certain coverage rate against different magnitudes of the attack $\epsilon$.

Theorems & Definitions (16)

• theorem 1
• Remark 1
• Proposition 1
• Remark 2
• proof
• Proposition 2
• proof
• Proposition 3
• proof
• Proposition 4