Title:	Deep Learning for Context-aware Motion Planning of Autonomus Vehicles
DNr:	Berzelius-2025-78
Project Type:	LiU Berzelius
Principal Investigator:	Paolo Falcone <paolo.falcone@chalmers.se>
Affiliation:	Chalmers tekniska högskola
Duration:	2025-08-04 – 2026-03-01
Classification:	20202
Keywords:

Abstract

This project provides GPU access to ongoing and planned research projects within context-aware motion planning of autonomous vehicles. This category of planning typically makes use of a prediction module that provides proposals for future trajectories of surrounding road users that is then passes into an optimization framework in order to compute a final plan for the autonomous vehicle. Recently, two types of approaches for implementing such planners have dominated the research field: 1) using deep neural networks for prediction combined with model predictive control algorithms for optimization and planning; and 2) using fully end-to-end optimized deep neural networks that predicts and plans jointly. Both approaches assume the capability to train neural networks at scale. We approach the subject from the perspective of the first method, which is often referred to as a hybrid planning method, with the goal of offering insights into how this class of planners can be understood and applied at scale within autonomous vehicle development. To this end, we aim to: 1) improve prediction algorithms through concepts such as pretraining, RL-finetuning, and distillation from foundation models; 2) condition prediction on proposed plans for the autonomous vehicle in order to capture the interactive property of human traffic behavior; and 3) package these improvements into methods that are computationally tractable whilst to some extent providing safety guarantees on the produced trajectory proposal. Finally, although context-aware planning often assumes perfect detection of surrounding road users, we see the need to benchmark our hybrid planners on raw sensor inputs by prepending well-established perception models in a two-stage training setup. This enables direct comparison with popular state-of-the-art end-to-end planners that operate on sensor inputs without intermediate abstraction.