Incremental Learning with Memory Regressors for Motion Prediction in Autonomous Racing

Yahan Yang, Souradeep Dutta, Kuk Jin Jang, Oleg Sokolsky, Insup Lee

Abstract: Cyber-physical systems (CPS) with learning-enabled components suffer from reduced performance under distribution shift. In this paper, we consider the problem of motion prediction within an autonomous racing setting. In such a setting, the ability to predict an adversaries' behavior is essential for safe and efficient planning. We propose a method using memories to detect anomalous input and incrementally learn a prediction model online, the ability to quickly adapt to unseen behaviors. In our experiments, we demonstrate the effectiveness of this approach in adapting to various motion prediction data collected in the F1Tenth-Gym environment, a simulator for autonomous racing. Our experiments show promising results, and achieve an improvement of 14% in mean squared error compared to a model without adaptation. In the future we would like to extend this to a more extensive evaluation of the ability of this core component to predict and act in an online racing setting.