A Physics-Informed CNN-LSTM for Speed and Heading-Rate Prediction from LiDAR Road Geometry Toward Safe Curve Negotiation

Anticipating the heading rate and speed that a curve will demand is a key capability for automated driving. This study presents a multi-task convolutional neural network–long short-term memory (CNN-LSTM) model that predicts the change of the vehicle heading rate, used as a steering proxy, and of its speed over horizons of 1, 2, and 3 s from a 2 s window of features. The inputs combine a preview of the road geometry extracted from light detection and ranging (LiDAR) point clouds (the lateral offset of the road at 10–40 m ahead), road banking and grade, and the vehicle speed and heading rate. A one-sided physics-informed penalty discourages predicted speeds above the maximum cornering speed of a banked curve with friction. The model was trained on PandaSet and evaluated on held-out sequences over five random sequence-wise splits, and without retraining on three sequences of the KITTI raw dataset, against a persistence baseline and an ablation without the physics term. On unseen PandaSet sequences, the model reduced the speed mean squared error with respect to persistence by 15.7%, 23.0%, and 27.0% at 1, 2, and 3 s, and the heading-rate error by 3.7%–16.4%. On KITTI, heading-rate error was on par with persistence, whereas speed error was 24–43% higher. The physics term was never active, since all speeds were well below the physical limit. The code is openly available.