MultiPath: Multiple Probabilistic Anchor Trajectory Hypotheses for Behavior Prediction

Predicting human behavior is a difficult and crucial task required for motion. It is challenging in large part due to the highly uncertain and-modal set of possible outcomes in real-world domains such as autonomous. Beyond single MAP trajectory prediction, obtaining an accurate distribution of the future is an area of active interest. We MultiPath, which leverages a fixed set of future state-sequence anchors correspond to modes of the trajectory distribution. At inference, our predicts a discrete distribution over the anchors and, for each anchor, offsets from anchor waypoints along with uncertainties, yielding a mixture at each time step. Our model is efficient, requiring only one inference pass to obtain multi-modal future distributions, and the is parametric, allowing compact communication and analytical queries. We show on several datasets that our model achieves more predictions, and compared to sampling baselines, does so with an order magnitude fewer trajectories.