Assessing internal models for faster learning of robotic assembly

This work investigates what makes a robotic assembly process “learnable” for the explicit purpose of improving the performance of that process. It has been observed that even stochastic search methods like Genetic Algorithms (GA) can benefit from advanced models of the assembly task. Models built from the results of random samplings of a parameter space have been used previously to predict the performances of parameter sequences not yet evaluated, but the question of what properties of the models actually benefit the optimization remained. A quantitative analysis algorithm is derived and tested on physical assemblies for validation. Results are provided that illustrate the efficacy of the analysis algorithm for prediction-based performance enhancement when such models are used.