Parameterized and Dynamic Gaussian Mixture Models for Robot Manipulation
This thesis compares task-parameterized Gaussian Mixture Models and Elastic Dynamical Systems for learning robot motions from demonstrations. It evaluates different fitting strategies and assesses how well the models adapt to changing task conditions.
.png)
Links
Overview
Robot learning has advanced quickly, with increasing use of autonomous systems in industry and daily life. Unlike traditional robots that repeat fixed motions, modern robots must adapt to new environments and task configurations, especially in human-robot settings where safety is critical.
Learning from Demonstrations is a common way to acquire robot policies from example trajectories. Gaussian Mixture Models are data-efficient and flexible for this purpose, but there are many variants and limited comparative evaluations. This thesis motivates a systematic comparison of GMM-based approaches for robust motion generalization.