PhD opportunities

Trajectory learning for the planning and control of autonomous vehicles

Area of expertise Real-time computer science, robotics,systems and control - Paris
Doctoral School SMI - Sciences des Mtiers de l'Ingnieur
Title Trajectory learning for the planning and control of autonomous vehicles
Supervisor M. Arnaud DE LA FORTELLE
Research unit Mathematics and Systems
CAOR - Centre de CAO et Robotique
Keywords Trajectory, Autonomous vehicles
Abstract In a complex operating environment, decision-making and control are increasingly based on hierarchical systems that allow the abstraction levels required for sophisticated reasoning to be reached while still maintaining the link to the elementary commands of Vehicles or to data from the sensors. Typically, it is a matter of recognizing the driving context (which requires a good level of abstraction) and of modulating both the perception and control algorithms: the environment will not be observed in the same way In the vicinity of a school in town at 16h that on a highway at midnight; Decisions will probably have little relevance in terms of planning and action.
Planning and control models are most often linked in order to ensure the proper execution of decisions. They are usually deterministic models. In order to limit the search space, planning often introduces aggregate maneuvers (change of file, vehicle follow-up, etc.) which allow the future trajectory to be broken down into a set of maneuvers that enable the driving intention to be realized Go to its destination). Different algorithms are implemented during each maneuver or change of maneuver.

The problem is how to predict the intentions of other users in order to adapt their behavior. The perception (detection / classification / tracking) is not the core of the thesis, although understanding (or even improving) some perceptual algorithms may be necessary. Similarly, the planning and control parts are important parts to understand how the thesis work will be exploited, but are already widely studied. The problem that arises is the following: from imperfect observations, how to deduce the intentions of other users, especially vehicles, in order to adjust its conduct. The work to be conducted covers many facets: clarification of the notion of intention (between maneuver decision, eg encoded by a homotopy class, and trajectory prediction), accessible data and data required for calculation. The most promising route would be to link learning to trajectory data (usually accessible from a fixed observatory) and on-board data (those derived from vehicle sensors and their fusion)
Funding Autre type de financement
Starting date October 1st 2017
Date of first publication July 12th 2017