MultiPath: Multiple Probabilistic Anchor Trajectory Hypotheses for\n Behavior Prediction

Abstract

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