Nocturne
A scalable driving benchmark for bringing multi-agent learning one step closer to the real world
Eugene Vinitsky*, Nathan Lichtlé*, Xiaomeng Yang*, Brandon Amos, Jakob Foerster
*These authors contributed equally to this work.
This page contains videos of Nocturne simulations, provided as a
supplementary material to the paper.
You can right-click on a video and "Show all controls" to change playback
speed or make a video fullscreen.
Scenes and partial-observability
These videos showcase the diversity of the scenes available in Nocturne,
and illustrate the partial-observability model we implement. All vehicles
in these videos are replayed from dataset trajectories.
Four-way stop
Straight road with merging vehicles
Unprotected turns with conflicting routes
Crowded parking lot
Diverse scenes
Partial-observability model
RL replays
These videos show replays of our trained reinforcement-learning
controllers on certain scenes, shedding light on different success and
failure modes. All non-stationary vehicles in these videos are RL-controlled
agents.
Conflicting goals at an intersection
The dark blue vehicle crashes into the pink vehicle which intersected
with its route while trying to turn left. All other vehicles make it to
their goals.
Truck failure
Here we observe both the orange and the green vehicles colliding with a
road edge (in green) after failing to turn early enough or hard enough.
Parking lot
All the vehicles reach their goals without crashing, despite being very
close to each other and having conflicting paths.
Straight road
Here again all the vehicles make it safely to their goals. Note the pink
vehicle that crosses a road edge is controlled by an expert (this is one
of the dataset limitations discussed in the paper of a road edge being
misclassified).
Narrow road
The lightgreen vehicle collides with a parked vehicle, as it failed to stay
on the narrow path leading to its goal.
Collision at an intersection
The pink and purple vehicles collide into each other, the other vehicles
reach their goals. Here many vehicles enter the intersection at once,
which is probably what led to the collision. On a more positive note, we
can observe the light purple vehicle on the right following the curve of
the road.
High-uncertainty intersection
All the vehicles reach their goals here. The red and blue vehicles
successfully took the bend.
Long straight road
Here we can observe some jiggling, due to high driving speed of the
vehicles and the discretization of the action space. In spite of that,
all vehicles reach their goals, except for the orange vehicle which drove
too close to a road edge and collided with it.
Parked collision
On a seemingly simple scenario, the green vehicle fails to turn and
crashes into a parked vehicle.
Waiting before merging
The purple vehicle successfully waited for the pink vehicle to pass
before speeding up.
Last updated on August 16th, 2022.