Using Roads and Sidewalks

Spawning on roads

On roads, route points, stored as carla.SumoNetworkRoutePoint objects, hold semantic information such the point's road, lane, and offset along lane. The SUMO network, stored as a carla.SumoNetwork, is used to traverse these route points spatially and topologically.

One way to spawn points is to first lookup the route point on the road nearest to an arbitrary position.

# Load SUMO network.
sumo_network = carla.SumoNetwork.load(PATH_TO_SUMO_NETWORK_FILE)

# Get arbitrary position.
position = carla.Vector(100, 100)

# Get nearest route point on SUMO network.
route_point = sumo_network.get_nearest_route_point(position)

Since the route point holds only semantic information and not the actual position, a translational query is required to determine the actual position to spawn the agent:

# Get route point position.
route_point_position = sumo_network.get_route_point_position(route_point)

# Get route point transform
route_point_transform = carla.Transform()
route_point_transform.x = route_point_position.x
route_point_transform.y = route_point_position.y
route_point_transform.z = 0.5 # Spawn at a height of 0.5 meters.

# Spawn actor. See CARLA's documentation.
world.spawn_actor(BLUEPRINT, route_point_transform)

Alternatively, you can use the carla.SegmentMap class to generate route points uniformly distributed along the lanes of the road. The carla.SegmentMap class is a data structure to work with line segments, optimized for uniform sampling of points over stored line segments.

# Get segments of SUMO network.
sumo_network_segments = sumo_network.create_segment_map()

# Randmly pick spawn point. Note that this returns a position directly, not a route point.
route_point_position = sumo_network_segments.rand_point()

It is also possible to bound the spawn segments to a certain region, for example a rectangular bounding box:

# Get segments of SUMO network.
sumo_network_segments = sumo_network.create_segment_map()

# Define bounding box.
bounds_min = carla.Vector2D(-50, -50)
bounds_max = carla.Vector2D(100, 300)
bounds_occupancy = carla.OccupancyMap(bounds_min, bounds_max)

# Calculate intersection of line segments with bounding box.
spawn_segments = sumo_network_segments.intersection(bounds_occupancy)

# Arbitrary polygon.
polygon = carla.OccupancyMap([carla.Vector2D(-50, -50),
    carla.Vector2D(-50, 100), carla.Vector2D(30, 200),
    carla.Vector2D(70, 70), carla.Vector2D(100, -70)])
# Arbitrary rectangle.
rectangle1 = carla.OccupancyMap(carla.Vector2D(-30, -30), carla.Vector2D(200, 300))
# Another arbitrary rectangle.
rectangle2 = carla.OccupancyMap(carla.Vector2D(-70, 30), carla.Vector2D(100, 200))
# Some combination of areas.
spawn_occupancy = polygon.union(rectangle1).difference(rectangle2)
# Crop line segments.
spawn_segments = sumo_network_segments.intersection(spawn_occupancy)

Navigation on roads

On roads, the SUMO network can be used to traverse the route points spatially and topologically.

The below example fetches the nearest route point given the agent's current position, and randomly selects from topologically possible next route points a set distance ahead.

# Get 2D position of actor.
location = actor.get_location()
position2d = carla.Vector2D(location.x, location.y)

# Lookup nearest route point.
route_point = sumo_network.get_nearest_route_point(position2d)

# Get all route points 1 meter ahead.
next_route_points = sumo_network.get_next_route_points(route_point, 1.0)

# Randomly select next route point.
next_route_point = random.choice(next_route_points)

Since the route point only holds semantic information and not the actual position, a translational query is required to determine the actual position:

# Get next route point's position.
position = sumo_network.get_route_point_position(next_route_point)

which can then be used for required tasks, such as for feedback to some steering controller for vehicles.

Spawning on sidewalks

On sidewalks, things work very similar to roads. Sidewalks also have route points, stored as carla.SidewalkRoutePoint objects, which hold semantic information such as the point's polygon, segment, and offset along segment. The sidewalk, stored as a carla.Sidewalk, is used to traverse these route points spatially and topologically.

One way to spawn points is to first lookup the route point on the road nearest to an arbitrary position.

# Load SUMO network.
sumo_network = carla.SumoNetwork.load(PATH_TO_SUMO_NETWORK_FILE)

# Calculate sidewalk 1.5 meters from road's mesh.
sumo_network_occupancy = sumo_network.create_occupancy_map()
sidewalk = sumo_network_occupancy.create_sidewalk(1.5)

# Get arbitrary position.
position = carla.Vector(100, 100)

# Get nearest route point on sidewalk.
route_point = sidewalk.get_nearest_route_point(position)

Since the route point holds only semantic information and not the actual position, a translational query is required to determine the actual position to spawn the agent:

# Get route point position.
route_point_position = sidewalk.get_route_point_position(route_point)

# Get route point transform
route_point_transform = carla.Transform()
route_point_transform.x = route_point_position.x
route_point_transform.y = route_point_position.y
route_point_transform.z = 0.5 # Spawn at a height of 0.5 meters.

# Spawn actor. See CARLA's documentation.
world.spawn_actor(BLUEPRINT, route_point_transform)

Alternatively, you can use the carla.SegmentMap class to generate route points uniformly distributed along the paths of the sidewalk. The carla.SegmentMap class is a data structure to work with line segments, optimized for uniform sampling of points over stored line segments.

# Get segments of sidewalk.
sidewalk_segments = sidewalk.create_segment_map()

# Randmly pick spawn point. Note that this returns a position directly, not a route point.
route_point_position = sidewalk_segments.rand_point()

It is also possible to bound the spawn segments to a certain region, for example a rectangular bounding box:

# Get segments of sidewalk.
sidewalk_segments = sidewalk.create_segment_map()

# Define bounding box.
bounds_min = carla.Vector2D(-50, -50)
bounds_max = carla.Vector2D(100, 300)
bounds_occupancy = carla.OccupancyMap(bounds_min, bounds_max)

# Calculate intersection of line segments with bounding box.
spawn_segments = sidewalk_segments.intersection(bounds_occupancy)

# Arbitrary polygon.
polygon = carla.OccupancyMap([carla.Vector2D(-50, -50),
    carla.Vector2D(-50, 100), carla.Vector2D(30, 200),
    carla.Vector2D(70, 70), carla.Vector2D(100, -70)])
# Arbitrary rectangle.
rectangle1 = carla.OccupancyMap(carla.Vector2D(-30, -30), carla.Vector2D(200, 300))
# Another arbitrary rectangle.
rectangle2 = carla.OccupancyMap(carla.Vector2D(-70, 30), carla.Vector2D(100, 200))
# Some combination of areas.
spawn_occupancy = polygon.union(rectangle1).difference(rectangle2)
# Crop line segments.
spawn_segments = sidewalk_segments.intersection(spawn_occupancy)

Navigating sidewalks

On sidewalks, the sidewalk can be used to traverse the route points spatially and topologically.

The below example fetches the nearest route point given the agent's current position, and selects the next route point a set distance anticlockwise along the route point's polygon.

# Get 2D position of actor.
location = actor.get_location()
position2d = carla.Vector2D(location.x, location.y)

# Lookup nearest route point.
route_point = sidewalk.get_nearest_route_point(position2d)

# Get route point 1 meter anticlockwise along route point's polygon.
next_route_point = sidewalk.get_next_route_point(route_point, 1.0)

To get the next route point clockwise (instead of anticlockwise) along the current route point's polygon, use the get_previous_route_point method instead:

# Get route point 1 meter clockwise along route point's polygon.
next_route_point = sidewalk.get_previous_route_point(route_point, 1.0)

You can also fetch the nearest point on an adjacent sidewalk polygon (i.e. on the other side of the road):

# Get nearest adjacent route point, at most 50 meters away.
# If such a route point exists, this returns None. a list of exactly one item.
next_route_point = sidewalk.get_adjacent_route_point(route_point, 50.0)

Since the route point only holds semantic information and not the actual position, a translational query is required to determine the actual position:

# Get next route point's position.
position = sidewalk.get_route_point_position(next_route_point)

which can then be used for required tasks, such as for feedback to some heading controller for pedestrians.