Session: CIE-23/24: AMS/SEIKM

Paper Number: 116329

116329 - A Physics-Informed Action Selection Framework for Robotic Heating 

Many processes require the surface temperature of the part to be maintained above a specific value. For example, metallurgical manufacturing processes that require targeted tempering, hardening, and annealing require localized heat treatment. This is also essential for local post-weld heat treatment processes that use localized welding to manufacture and repair large parts like turbine blades. Localized heating of materials is also used in applications to make materials temporarily more pliable. Automated prepreg carbon fiber composite applications use localized heating to activate the resin in the sheet to conform it to the composite part mold. Thermal curing of adhesives and composite sheets also has similar controlled temperature heating requirements. Applications like localized plastic welding and spot curing of thermal adhesives require localized heating and curing of bonded assemblies. The heating process needs to be performed efficiently. Overheating can cause serious damage. On the other hand, not achieving the desired temperature interferes with the process. Usually, the objective is to achieve the desired surface temperature in the fastest possible time without causing overheating. We believe that the robot can be used to control the position of the heat source and maintain the desired temperature. Automating this process can lead to a consistent process and eliminates the possibility of an error. Robots are being considered for performing external heating of components in manufacturing applications. This paper presents a physics-aware action selection policy that employs forward simulation with a branch and bound search to efficiently determine the best action sequence to position the heating tool. We take inspiration from physics-informed machine learning and present a parameter learning graph-based modeling framework that enables the robot to predict the temperature evolution of the surface of interest with respect to time. We also present a state transition model to describe how the thermal characteristics of the system change based on the heating tool's position. We demonstrate our proposed robotic heating action selection approach for the composite layup process on an industrial tool. This work demonstrates the usefulness of physics-inspired machine learning in a real-world application.

Presenting Author: Satyandra Gupta University of Southern California

Presenting Author Biography: SK Gupta is a Professor at the University of Southern California.

Authors:

Neel Dhanaraj University of Southern California
Omey Manyar University of Southern California
Vihan Krishnan University of Southern California
Satyandra Gupta University of Southern California