Department of Electronic and Computer Engineering Seminar - Congestion-Aware Distribution Transport and Few-Step Control for Traffic Flow and Robot Swarms

The motion of large groups of vehicles or robots is often better described at the level of distributions than at the level of individual agents. This macroscopic viewpoint is natural in traffic flow, swarm coordination, and other large-scale systems where one is interested in how a population moves, spreads, and redistributes over space and time. A central difficulty in such problems is congestion: when density becomes large, motion slows down, throughput saturates, and not every transport plan is physically realizable.

In this talk, I will present a congestion-aware dynamic optimal transport framework that incorporates this effect directly into the transport model through a fundamental-diagram constraint. This leads to distribution transport plans that are not only efficient in effort, but also compatible with density-dependent capacity limits. While the idea is motivated by classical traffic flow theory, it also has a clear interpretation in robotics, where robot swarms moving through narrow passages, doors, corridors, or shared workspaces face similar congestion effects at the collective level.

I will then discuss how this modeling viewpoint connects to control. In particular, I will describe how generative-model ideas, through a MeanFlow-type framework, can be used to learn few-step policies for steering distributions efficiently toward desired targets. This provides a practical mechanism for fast distribution control while remaining consistent with the underlying transport structure. Taken together, these ideas suggest a unified perspective in which congestion-aware transport provides the macroscopic model, and generative learning provides an efficient tool for control, with applications to both traffic flow and robot swarms.