Physics-informed deep learning for robust trajectory prediction in automated driving.

Planning a safe, comfortable, and energy-efficient ego-vehicle trajectory is fundamental for automated driving. Efficiency reduces emissions, while safety protects occupants and vulnerable road users. Purely data-driven models, however, struggle with robustness and out-of-distribution scenarios, especially for rare edge cases which are underrepresented in the training data. If the data is furthermore affected by sensor noise, the model performance can suffer drastically. Physics-Informed Neural Networks incorporate expert knowledge in form of physical laws or governing equations, enhancing generalisation and transparency compared to standard data-driven models. Only a few studies have applied such networks to the use case automated driving because of lack of reliable physical models for the complex human-driving patterns. This paper investigates the benefit of injecting a physical model, which describes the motion patterns only to a certain extent. It is demonstrated that it still enables a highly accurate extrapolation to OOD scenarios and yields robust performance even with limited data. The root mean squared error on the predicted trajectories is reduced up to a factor of 10.