In various examples, a method includes computing a current keyframe, the current keyframe being representative of an area around an autonomous vehicle at a current time based on map data. The method includes transforming a preceding keyframe to a coordinate frame of the autonomous vehicle at a first time prior to completing computation of the current keyframe to generate a first world model frame. The method includes transforming the preceding keyframe to the coordinate frame of the autonomous vehicle at a second time after the first time and prior to completing computation of the current keyframe to generate a second world model frame.

Aspects of the disclosure relate to generating map data. For instance, data generated by a perception system of a vehicle may be received. This data corresponds to a plurality of observations including observed positions of a passenger of the vehicle as the passenger approached the vehicle at a first location. The data may be used to determine an observed distance traveled by a passenger to reach a vehicle. A road edge distance between an observed position of an observation of the plurality of observations and a nearest road edge to the observed position may be determined. An inconvenience value for the first location may be determined using the observed distance and the road edge distance. The map data is then generated using the inconvenience value.

Aspects of the disclosure provide for the generation of a service area map for autonomous vehicles. For instance, graph nodes of a road network may be iterated through in order to identify a set of reachable graph nodes based on a set of routing parameters that define driving limits for the autonomous vehicles. The road network may include the graph nodes as well as edges connecting ones of the graph nodes. A set of S2 cells may be identified based on the set of reachable graph nodes. Vertices of each S2 cell of the set of S2 cells may be determined based on whether each S2 cell of the set of S2 cells is occupied by any of the graph nodes of the set of reachable graph nodes. Contours through cells may be drawn based on the scores. The service area map may be generated using the contours.

A method, apparatus and computer program product are provided for learning to generate maps from raw geospatial observations from sensors traveling within an environment. Methods may include: receiving a plurality of sequences of geospatial observations from discrete trajectories; aligning and concatenating the trajectories to generate concatenated, aligned geospatial observations; performing attentional clustering on the concatenated, aligned geospatial observations to obtain a set of entities; processing the set of entities using one or more Set Transformers; generating a preliminary feature set of map object geometries; determining a number of output entities from each of the one or more output heads based on statistics from the plurality of sequences of geospatial observations from the discrete trajectories and a count of the discrete trajectories; and outputting a feature set, the feature set including a number of map object geometries corresponding to a respective number of output entities for the respective output head.

Systems and methods are provided for generating a crowd-sourced map for use in vehicle navigation. In one implementation, a system may include at least one processor configured to receive drive information collected from vehicles that traversed a junction; aggregate the received drive information to determine positions of traffic lights and spline representations for drivable paths; input the determined positions and the spline representations to a trained model configured to generate a traffic light relevancy mapping indicating a traffic light relevancy for traffic light to drivable path pairs of the junction; input an observed vehicle behavior to the at least one trained model to generate an updated traffic light relevancy mapping; store in the crowd-sourced map the indicators of traffic light relevancy for the traffic light to drivable path pairs; and transmit the crowd-sourced map to a vehicle for use in navigating the road segment.

A method, apparatus and computer program product are provided for learning to generate maps from raw geospatial observations from sensors traveling within an environment. Methods may include: receiving a plurality of sequences of geospatial observations from discrete trajectories; refining a summary entity set representation through iteration over discrete trajectories; determining a drive offset for each of the discrete trajectories based on a comparison of the summary entity set representation to a drive entity set for each of the discrete trajectories; aligning the discrete trajectories to generate aligned geospatial observations based, at least in part, on the drive offset for a respective discrete trajectory; concatenating the aligned geospatial observations; processing the concatenated, aligned geospatial observations using at least one Set Transformer; and generating, from the at least one Set Transformer, map geometries including objects from the geospatial observations.

A method, apparatus and computer program product are provided for learning to generate maps from raw geospatial observations from sensors traveling within an environment. Methods may include: processing geospatial observations from discrete trajectories through an iterative attention model incorporating a Gated Recurrent Unit gating pattern to obtain a feature summary; determining a drive offset for each of the discrete trajectories based on the feature summary; aligning the discrete trajectories to generate aligned geospatial observations based, at least in part, on the drive offset for a respective discrete trajectory; concatenating the aligned geospatial observations; processing the concatenated, aligned geospatial observations using at least one Set Transformer; generating, from the at least one Set Transformer, map geometries including objects from the geospatial observations; and providing for at least one of navigational assistance or at least semi-autonomous vehicle control based on the map geometries.

A method for updating a localization of an autonomous vehicle includes receiving perception input data from a sensor of the autonomous vehicle and receiving map data including road lane information in a vicinity of the autonomous vehicle. The method includes processing the perception input data to extract perceived road lane information including a perceived x position, a perceived y position, a perceived z position, a perceived lane type, a perceived lane color, and a perceived lane curvature and processing the map data to extract map road lane information including a map x position, a map y position, a map z position, a map lane type, a map lane color, and a map lane curvature. The method includes calculating a transformation matrix from the perceived road lane information and the map road lane information and updating the map data and a localization of the vehicle based on the transformation matrix.

Techniques associated with generating and maintaining sparse geographic and map data. In some cases, the system may maintain a factor graph comprising a plurality of nodes. In some cases, the nodes may comprise pose data and sensor data associated with an autonomous vehicle at the geographic position represented by the node. The nodes may be linked based on shared trajectories and shared sensor data.

A method, apparatus and computer program product are provided for generating structured trajectories based on probe data while maintaining privacy and user information. Methods may include: receiving a plurality of sequences of probe data points from a plurality of probe apparatuses; identifying splitting points in each of the plurality of sequences of probe data points; identifying legs of the plurality of sequences of probe data points between pairs of splitting points; grouping legs within a predefined degree of similarity into bunches of legs; performing a guided search of a solution space containing the bunches of legs by performing successive mutations on candidate solutions in the solution space to identify a solution satisfying a fitness metric threshold; and identifying, from the solution satisfying a fitness metric threshold, a road network.