Predicting how other traffic will move

Development of a hybrid method for the motion prediction of other traffic participants. Submitted 15 October 2021, six months’ work, done inside MAN’s Hamburg TruckPilot project.

Motion prediction is the question of what the car beside you is about to do. An automated vehicle cannot plan without an answer, and the answer is unavoidably a guess — you are inferring intent from geometry and history, about a driver you cannot see.

There are two classical ways to guess and each fails in the opposite direction. A physics-based model propagates what the object is doing now — velocity, yaw rate, acceleration — and is excellent for the next second and useless after three, because it has no notion that the driver might want something. A manoeuvre- or learning-based model predicts intent from patterns, and is much better further out, but it will confidently produce a trajectory that violates the vehicle’s own dynamics if nothing stops it.

A hybrid method is the attempt to have both: the near horizon governed by physics, the far horizon governed by intent, and a principled way of handing over between them.

The part that reads differently now

The thesis was set inside Hamburg TruckPilot, MAN’s autonomous truck project. MAN belongs to the TRATON GROUP.

Four years later I was leading a feature workstream on the TRATON Level 4 programme at Plus. I did not plan that, and I did not notice the symmetry until I wrote this page — but the first serious autonomy work I did as a student and the partner programme I later helped put on public roads turn out to belong to the same company.