Results

Below we display some videos and results from our work and publications.

Shockwave Dissipation

In this work, inspired by the Sugiyama experiment (cite) demonstrating spontaneous formation of traffic shockwaves, and experiments demonstrating the ability of AVs to suppress shockwaves (cite), we investigated the ability of reinforcement learning to train an optimal shockwave dissipating controller.

By training on rings of randomly varied size, and using a neural net with memory, we were able to learn a controller that both was optimal and generalized outside of the training distribution.

The figure shows a diagram of the experiment, in which a single AV tries to stabilize an unstable ring.

ring-scenario
Bottleneck control design
(A). Average system velocity before and after the controller is turned on. (B). Space-time diagram of the evolution of shockwaves in the system. (C). A ring control experiment done with a single autonomous vehicle (cite DAN). (D) The 2008 Sugiyama experiment [cite]
ring-scenario
Comparisons of average system velocity under different control strategies, with the theoretical optimum as a benchmark. The GRU (memory) controller both reached the optimal on the training set, shown in white, and generalized out of distribution to the densities shown in grey.

Intersection control

In this work, we demonstrated the ability of a single autonomous vehicle to control the relative spacing of vehicles following behind it to create an optimal merge.

As can be seen in the videos, at low penetration rates the autonomous vehicle bunches the vehicles to avoid the intersection, whereas at high penetrations the optimal solution becomes well timed merges.

Intersection-scenario.
Diagram of the intersection experiment. 14 vehicles trying to cross an intersection.
Intersection-scenario.
Space-time diagram of the intersection experiment. 14 vehicles trying to cross an intersection with no control.
Intersection-scenario.
Space-time diagram of the intersection experiment. One of the vehicles is autonomous.
Intersection-scenario.
Space-time diagram of the intersection experiment. All the vehicles are autonomous.

Videos representing the depicted space-time diagrams are below. Highlighted vehicles are autonomous.


On-ramp shockwave dissipation

In this work, we demonstrate the ability of shockwave dissipation, previously demonstrated on a ring to scale to an actual highway situation. Vehicles travelling on a highway are purturbed by an aggressive on-ramp merge that the autonomous vehicle needs to dissipate.

As can be seen in the videos, the autonomous vehicle learns to slow its following vehicles to either avoid the perturbation or smooth its effects.

Bottleneck control design
Architecture of the merge control; AVs can only sense the vehicle in front and behind them.
Bottleneck control design
In the limit of a large ring, there is an equivalence between an on-ramp merge and a perturbation in a ring.
Bottleneck control design
Space-time diagram of highway vehicles w/ and w/out the inclusion of AVs.
Bottleneck control design
Control structure of the bottleneck
Bottleneck control design
Control structure of the bottleneck

Bottleneck control

Inspired by the rapid decrease in lanes on the San Francisco-Oakland Bay Bridge, we study a bottleneck that merges from four lanes down to two to one.

We demonstrate that the AVs are able to learn a strategy that increases the effective outflow at high inflows, and performs competitively with ramp metering.

Bottleneck control design
Control structure of the bottleneck. Scale of segments are distorted for visualization.
Bottleneck control design
Without control, congestion rapidly forms in the bottleneck.
Bottleneck control design
Control structure of the bottleneck; at high inflows the outflow is improved by 25%.
Bottleneck control design
Comparison of inflow, outflow curves for AV control vs. ramp metering. At high inflows they perform comparably.

Impact of sorting

Bottleneck control design
Control structure of the bottleneck
Bottleneck control design
Effect of sorting on 6 rl and 38 human
Bottleneck control design
Effect of sorting on 44 RL vehicles