ANALYSIS OF CONTINUOUS LEARNING MODELS FOR TRAJECTORY REPRESENTATION

 Trajectory planning is a field with widespread utility, and imitation learning pipelines
show promise as an accessible training method for trajectory planning. MPNet is the state
of the art for imitation learning with respect to success rates. MPNet has two general
components to its runtime: a neural network predicts the location of the next anchor point in
a trajectory, and then planning infrastructure applies sampling-based techniques to produce
near-optimal, collision-less paths. This distinction between the two parts of MPNet prompts
investigation into the role of the neural architectures in the Neural Motion Planning pipeline,
to discover where improvements can be made. This thesis seeks to explore the importance
of neural architecture choice by removing the planning structures, and comparing MPNet’s
feedforward anchor point predictor with that of a continuous model trained to output a
continuous trajectory from start to goal. A new state of the art model in continuous learning
is the Neural Flow model. As a continuous model, it possess a low standard deviation runtime
which can be properly leveraged in the absence of planning infrastructure. Neural Flows also
output smooth, continuous trajectory curves that serve to reduce noisy path outputs in the
absence of lazy vertex contraction. This project analyzes the performance of MPNet, Resnet
Flow, and Coupling Flow models when sampling-based planning tools such as dropout, lazy
vertex contraction, and replanning are removed. Each neural planner is trained end-to-end in
an imitation learning pipeline utilizing a simple feedforward encoder, a CNN-based encoder,
and a Pointnet encoder to encode the environment, for purposes of comparison. Results
indicate that performance is competitive, with Neural Flows slightly outperforming MPNet’s
success rates on our reduced dataset in Simple2D, and being slighty outperformed by MPNet
with respect to collision penetration distance in our UR5 Cubby test suite. These results
indicate that continuous models can compete with the performance of anchor point predictor
models when sampling-based planning techniques are not applied. Neural Flow models also
have other benefits that anchor point predictors do not, like continuity guarantees, the ability
to select a proportional location in a trajectory to output, and smoothness.