Abstract: Techniques for training a codebook usable by a machine learned model to predict an object trajectory or scene data are described herein. For example, the techniques may include generating tokens representing discrete object behavior into a machine learned model that outputs a sequence of tokens that is usable by another machine learned model to generate the object trajectory (e.g., position data, velocity data, acceleration data, etc.) or the scene data associated with the environment. The object trajectory can be sent to a vehicle computing device for consideration during vehicle planning, which may include simulation.

Abstract: Techniques for predicting an object trajectory or scene information are described herein. For example, the techniques may include inputting tokens representing discrete behavior into a machine learned model. The machine learned model may output a sequence of tokens that is usable by another machine learned model to generate an object trajectory (e.g., position data, velocity data, acceleration data, etc.) for one or more objects in the environment. The object trajectory can be sent to a vehicle computing device for consideration during vehicle planning, which may include simulation.

Abstract: Techniques for predicting an object trajectory or scene information are described herein. For example, the techniques may include inputting latent variable data into a machine learned model. The machine learned model may output an object trajectory (e.g., position data, velocity data, acceleration data, etc.) for one or more objects in the environment based on the latent variable data. The object trajectory can be sent to a vehicle computing device for consideration during vehicle planning, which may include simulation.

Abstract: Techniques for increasing performance of machine-learned models while conserving computational resources generally required by ensemble machine-learning methods are described herein. The techniques may include determining multiple views of a scene that is to be input into a machine-learned model. In some examples, a scene data input may be rotated by 90, 180, and 270 degrees to generate four scene inputs (e.g., 0-, 90-, 180-, and 270-degree rotated inputs) that can be passed through the machine-learned model and the results per scene can be aggregated to determine a final prediction/decision. Similarly, scene inputs may be shifted, reflected, translated, and/or the like before being input into the machine-learned model. The predictions may be associated with one or more objects in the environment that are represented in the scenes.

Abstract: Techniques for improving operational decisions of an autonomous vehicle are discussed herein. In some cases, a system may generate reference graphs associated with a route of the autonomous vehicle. Such reference graphs can comprise precomputed feature vectors based on grid regions and/or lane segments. The feature vectors are usable to determine scene context data associated with static objects to reduce computational expenses and compute time.

Abstract: Techniques for selectively offloading data that is computed by a first processing unit during training of an artificial neural network onto memory associated with a second processing unit and transferring the data back to the first processing unit when the data is needed for further processing are described herein. For example, the first processing unit may compute activations for operations associated with forward propagation. During the forward propagation, one or more of the activations may be transferred to a second processing unit for storage. Then, during backpropagation for the artificial neural network, the activations may be transferred back to the first processing unit as needed to compute gradients.

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 first predicted position of the object. The first predicted position may be determined to be outside of a bounded area of the environment. Based on this determination, a second predicted position of the object may be determined using map data associated with 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 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 training a machine learning model are described herein. For example, the techniques may include implementing a cross batch normalization layer that generates a cross batch normalization layer output based on a first layer output during training of the neural network. The training may be based on a local batch of training examples of a global batch including the local batch and at least one remote batch of training examples. The cross batch normalization layer output may include normalized components of the first layer output determined based on global normalization statistics for the global batch. Such techniques may be used to train a neural network over distributed machines by synchronizing batches between such machines.

Abstract: Techniques for training a machine learning model are described herein. For example, the techniques may include implementing a cross batch normalization layer that generates a cross batch normalization layer output based on a first layer output during training of the neural network. The training may be based on a local batch of training examples of a global batch including the local batch and at least one remote batch of training examples. The cross batch normalization layer output may include normalized components of the first layer output determined based on global normalization statistics for the global batch. Such techniques may be used to train a neural network over distributed machines by synchronizing batches between such machines.