Predict pedestrians' trajectories by estimating their motion planners

Abstract

Predicting an agent's (in this case human's) trajectory is a challenging, yet crucial step in robotic research. Humans have "internal planners" that govern their paths, which allow them to dynamically plan their trajectories. Robots can safely navigate and interact with people by simulating their planners to predict their future trajectories. To allow robots to predict the agent's future trajectory, many of the current approaches fit statistical models through a set of collected human trajectories and use that fitted model to predict future trajectories. These current approaches require a large variety of training data, and they break when the trajectory distribution changes due to the changes in the environment. Human solves this problem by simulating other agents' path planning behavior to accurately predict their future trajectories. This technique is called the simulation theory of the mind. Implementing this technique allows the robot to perform robustly in dynamic, changing environments. The aim of this project is to combine the Kernel Density Estimator (KDE), Mixture Density Network (MDN) with Long Short-Term Memory Network (LSTM) to produce a KDE-LSTM-MDN Network that can learn the complex path planning behavior of an agent, thus improving the sampling process of a Rapidly Exploring Random Tree (RRT*) to estimate a generalized planner that can be used to predict future trajectories of people. We evaluate the performance of our system in simulations.