Abstract: A machine-learned architecture may predict multiple paths that an object could take in the future without regard to time at which the object may occupy positions identified by one of those paths. These time-invariant paths may be used by an autonomous vehicle to filter detected objects by relevance to an autonomous vehicle's plans, improve prediction of an object's reaction to a vehicle candidate trajectory, determine right-of-way between object(s) and the autonomous vehicle, match detected objects to lanes, and/or improve prediction of odd or out-of-turn object behavior of an object.

Abstract: A machine-learned architecture may predict a set of spatially-diverse paths that an object may take in the future. The paths generated by this architecture may be time-invariant (e.g., not identifying a time at which the object may occupy a position along one of these paths) but can be used by a second machine-learned model to predict progress in time along these paths. This segregation of the spatial paths and progress in time along the paths improves the accuracy of the ultimate prediction and better captures rare object behavior.

Abstract: Techniques for identifying road segments associated with an object trajectory are discussed herein. A computing device can implement a model that extracts points from an object trajectory and identifies road segments along a path between the extracted points. The model can determine a set of road segments for the path based on searching a graph. The graph can comprise nodes to represent different road segments, and a graph search algorithm can output road segments representing a shortest path between the nodes associated with the extracted points. The road segments can be used by a vehicle computing device for predicting vehicle actions to control a vehicle.

Abstract: Techniques for determining unified futures of objects in an environment are discussed herein. Techniques may include determining a first feature associated with an object in an environment and a second feature associated with the environment and based on a position of the object in the environment, updating a graph neural network (GNN) to encode the first feature and second feature into a graph node representing the object and encode relative positions of additional objects in the environment into one or more edges attached to the node. The GNN may be decoded to determine a distribution of predicted positions for the object in the future that meet a criterion, allowing for more efficient sampling. A predicted position of the object in the future may be determined by sampling from the distribution.

Abstract: Techniques for determining unified futures of objects in an environment are discussed herein. Techniques may include determining a first feature associated with an object in an environment and a second feature associated with the environment and based on a position of the object in the environment, updating a graph neural network (GNN) to encode the first feature and second feature into a graph node representing the object and encode relative positions of additional objects in the environment into one or more edges attached to the node. The GNN may be decoded to determine a distribution of predicted positions for the object in the future. A predicted position of the object at a subsequent timestep may be determined by sampling from the distribution of predicted positions according to various sampling strategies. Alternatively, the predicted position of the object may be overwritten using a candidate position of the object.

Abstract: Techniques for determining unified futures of objects in an environment are discussed herein. Techniques may include determining a first feature associated with an object in an environment and a second feature associated with the environment and based on a position of the object in the environment, updating a graph neural network (GNN) to encode the first feature and second feature into a graph node representing the object and encode relative positions of additional objects in the environment into one or more edges attached to the node. The GNN may be decoded to determine a predicted position of the object at a subsequent timestep. Further, a predicted trajectory of the object may be determined using predicted positions of the object at various timesteps.

Abstract: Techniques for generating simulations to evaluate an update to a controller. The controller may be configured to control one or more functionalities of an autonomous and/or a semi-autonomous vehicle. A simulation computing system may receive a request to evaluate a first controller. The simulation computing system may generate a simulation based on data associated with a previous operation of the vehicle in an environment, the previous operation being controlled by a second controller (e.g., standard for evaluation, control version, etc.). The simulation computing device may cause the first controller to control a simulated vehicle in the simulation and may determine whether to validate the update to the controller based on a difference between first metrics associated with a control of the simulated vehicle by the first controller and second metrics associated with a control of the vehicle by the second controller.

Abstract: Techniques for determining unified futures of objects in an environment are discussed herein. Techniques may include determining a first feature associated with an object in an environment and a second feature associated with the environment and based on a position of the object in the environment, updating a graph neural network (GNN) to encode the first feature and second feature into a graph node representing the object and encode relative positions of additional objects in the environment into one or more edges attached to the node. The GNN may be decoded to determine a predicted position of the object at a subsequent timestep. Further, a predicted trajectory of the object may be determined using predicted positions of the object at various timesteps.

Abstract: Techniques for determining unified futures of objects in an environment are discussed herein. Techniques may include determining a first feature associated with an object in an environment and a second feature associated with the environment and based on a position of the object in the environment, updating a graph neural network (GNN) to encode the first feature and second feature into a graph node representing the object and encode relative positions of additional objects in the environment into one or more edges attached to the node. The GNN may be decoded to determine a distribution of predicted positions for the object in the future that meet a criterion, allowing for more efficient sampling. A predicted position of the object in the future may be determined by sampling from the distribution.