Predicting Agent Behavior by Estimating Motion Planners

Abstract

To navigate the world safely, autonomous agents must predict the future actions of the other agents in the world. We propose a method to estimate the future positions of other agents by using a sample-based planner that samples from a distribution trained on observed vehicle trajectories. The sample-based planner used is Rapidly-exploring Random Tree, and trained distribution is the combination of two networks: a Long Short-Term Memory network and a Mixture Density Network. We compared this trained model to several baselines. The proposed method performs better than the curve fitting and Long Short-Term Memory baselines, but worse than the other two baselines. The methods with a path planner perform better than the baselines without, which may mean that using a path planner to complete the problem of vehicle trajectory prediction may prove beneficial in future works.