Note: This content is accessible to all versions of every browser. However, this browser does not seem to support current Web standards, preventing the display of our site's design details.


Supervised learning of the safe set in autonomous driving



Alexander Liniger

In previous work we investigated the use and benefits of safe sets for the control of autonomous race cars. More precisely we used the viability kernel to guarantee recursive feasibility of planned trajectories by only exploring viable and therefore safe actions. However, to apply the method the viability kernel needs first do be computed, which needs the exact knowledge of the constraint set as well as the dynamics. In autonomous driving where compared to autonomous racing, not a predefined race track but a general road network is of interest, it is impossible to compute the viability kernel. However, as roads are normally a sequence of straight segments connected with curves, the goal of this thesis is to investigate, if it is possible to learn the safe set of general roads by only looking at a library of curves.

Therefore, in a first phase a library of viability kernels for curves with different radius and angle has to be generated. To do so the existing viability computation framework has to be modified and suited boundary conditions for the viability computation have to be determined. In a second phase, based on this library ideas form supervised learning, e.g., support vector machines, should be used to learn which states are safe. The learned classifier should then be tested with yet unseen curves.

Weitere Informationen


John Lygeros

Art der Arbeit:
Anzahl StudentInnen: 1
Status: done
Projektstart: March 2017
Semester: FS17