Date of Award

12-2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Aerospace, Physics, and Space Sciences

First Advisor

Madhur Tiwari

Second Advisor

Vivek Muralidharan

Third Advisor

Xianqi Li

Fourth Advisor

Donald Platt

Abstract

Imitation learning offers a promising approach in developing near-optimal onboard policies for nonlinear systems, by learning from computationally expensive trajectory planners during offline training. Existing State-of-the-art methods commonly learn deterministic mappings from observations to actions by training fast neural networks to mimic the expert demonstrations, which can limit their ability to capture the action distribution, when the policy is trained on multiple possible expert commands to reach the same target, and this may lead to unsafe scenarios in the obstacle environments. To address these challenges, this study explores Diffusion Models (DMs) as a feedback controller for quadcopters that learns the distribution of action sequences conditioned on the observation. Results from the simulation suggests that the trained DM Policy captures multimodal action distributions while avoiding obstacles. Hardware tests validates that the DM Policy mimics the expert demonstrations. Furthermore, the role of noise schedules in the DM framework, incorporating obstacle information into conditioning, and comparison with the behaviour cloning policy are discussed.

Download

Share

COinS