Training and executing navigation models for autonomous vehicles

Abstract

Techniques are discussed herein for training and executing movement and navigation models for controlling autonomous vehicles in simulations and/or in real-world physical environments. In some examples, a model training component may receive object movement data, such as velocities and/or steering angles, from vehicle logs or other ground truth data sources. The movement data may be discretized into labeled bins, and a smoothing function (e.g., a Gaussian quadrature or other weight sum function) centered on the ground truth data may be applied to the labeled bins. During model training, the model outputs can be compared to the smoothed object movement data using a loss function (e.g., a cross-entropy loss function), to reduce the magnitude of the loss penalties for incorrect predictions that are closer to the ground truth data. As a result, models trained using the techniques described herein may provide improved object movement predictions and/or navigation performance for controlling autonomous vehicles.