Model-Free Generative Sampling of Continuum-Robot Configurations for Vision-Based Motion Planning
Image base Motion Planning needs either a dense roadmap of valid configurations or a way to sample valid robot configuration in the image space. Analytical models of soft robots exist, but they might not be accurate due to material properties, discrepancies through construction. They also usually do not take into account the dynamic forces on the robot and might not be available for novel robots. The set of valid configurations needed for motion planning can be acquired by moving the robot to collect data or learned efficiently through observations. Data efficiency is a central concern here as it translates to long data collection hours.
This thesis attempt to understand if the compositional structure of a multi-module origami robot can be used to make this problem tractable. We compare three families of generative models- flat (which treat the configuration as a single vector), graph structured decoders (which attempt to learn the relationship between key points) and a compositional network which trains a generative model on only 1 module and chains modules together using SE(2) transformations, building the structure rather than learning it.
The models are compared in simulation, where curve-fitting projection errors serve as a measure of sample validity, and performance is tracked as training set shrinks. Preliminary results suggest that graph-structured decoders do not learn about the structure efficiently which motivates a closer look at the compositional model. The best model is then deployed on a multi-module origami arm to build a roadmap. The paths generated are then executed using visual servoing.
Advisor: Professor Berk Calli
Committee: Professor Vincent Aloi, Professor Griffn Tabor
Zoom link: https://wpi.zoom.us/j/91655971603