Safe, Out-of-Distribution-Adaptive MPC with Conformalized Neural Network Ensembles
Authors:
Jose Leopoldo Contreras,
Ola Shorinwa,
Mac Schwager
Abstract:
We present SODA-MPC, a Safe, Out-of-Distribution-Adaptive Model Predictive Control algorithm, which uses an ensemble of learned models for prediction, with a runtime monitor to flag unreliable out-of-distribution (OOD) predictions. When an OOD situation is detected, SODA-MPC triggers a safe fallback control strategy based on reachability, yielding a control framework that achieves the high perform…
▽ More
We present SODA-MPC, a Safe, Out-of-Distribution-Adaptive Model Predictive Control algorithm, which uses an ensemble of learned models for prediction, with a runtime monitor to flag unreliable out-of-distribution (OOD) predictions. When an OOD situation is detected, SODA-MPC triggers a safe fallback control strategy based on reachability, yielding a control framework that achieves the high performance of learning-based models while preserving the safety of reachability-based control. We demonstrate the method in the context of an autonomous vehicle, driving among dynamic pedestrians, where SODA-MPC uses a neural network ensemble for pedestrian prediction. We calibrate the OOD signal using conformal prediction to derive an OOD detector with probabilistic guarantees on the false-positive rate, given a user-specified confidence level. During in-distribution operation, the MPC controller avoids collisions with a pedestrian based on the trajectory predicted by the mean of the ensemble. When OOD conditions are detected, the MPC switches to a reachability-based controller to avoid collisions with the reachable set of the pedestrian assuming a maximum pedestrian speed, to guarantee safety under the worst-case actions of the pedestrian. We verify SODA-MPC in extensive autonomous driving simulations in a pedestrian-crossing scenario. Our model ensemble is trained and calibrated with real pedestrian data, showing that our OOD detector obtains the desired accuracy rate within a theoretically-predicted range. We empirically show improved safety and improved task completion compared with two state-of-the-art MPC methods that also use conformal prediction, but without OOD adaptation. Further, we demonstrate the effectiveness of our method with the large-scale multi-agent predictor Trajectron++, using large-scale traffic data from the nuScenes dataset for training and calibration.
△ Less
Submitted 4 June, 2025; v1 submitted 4 June, 2024;
originally announced June 2024.
