Selim, Mahmoud (KTH Royal Institute of Technology, Scania CV AV), Bhat, Sriharsha (KTH Royal Institute of Technology; Scania CV AB), Johansson, Karl H. (KTH Royal Institute of Technology)

Motion Planning Using Physics-Informed LSTMs for Autonomous Driving

Scheduled for presentation during the Invited Session "Learning-powered and Knowledge-driven Autonomous Driving II" (ThBT1), Thursday, September 26, 2024, 15:50−16:10, Salon 1

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 8, 2024

Abstract

Developing accurate models for the behavior of heavy-duty vehicles such as trucks and buses is crucial for ensuring their safe navigation. It is important that these models accurately reflect the vehicle’s performance across different weather conditions, road types, and cargo loads. This paper presents the use of Physics-Informed Long Short-Term Memory (PI-LSTM) networks as dynamic models tailored for samplingbased motion planning, which is relevant for the navigation of autonomous vehicles. By combining the LSTM’s capability to predict nonlinear vehicle dynamics with physics-based constraints incorporated into the loss function, our planner will be able to generate motion plans that are not only more dynamically accurate but also efficient and capable of full parallel execution on a GPU to significantly enhance planning speed. Evaluating our model using real data from vehicle tests on different road surfaces and driving maneuvers, we see that PI-LSTMs capture vehicle behavior with a significantly lower error than traditional modelling techniques.