Optimization of Machine Learning Models Applied to Robot Localization in the RobotAtFactory 4.0 Competition

Several approaches have been developed over time aiming to improve the localization aspects, especially in mobile robotics. Besides the more traditional techniques, mainly based on analytical models, artificial intelligence has emerged as an interesting alternative. The current study proposes to explore the machine learning model structure optimization for pose estimation, using the RobotAtFactory 4.0 competition as the main context. Using a Bayesian Optimization-based framework, the parameters of a Multi-Layer Perceptron (MLP) model, trained to estimate the components of the 2D pose (x, y, and \(\theta \) ) of the robot were optimized in four different scenarios of the same context. The results obtained showed a quality improvement of up to 60% on the estimation when compared with the modes without any optimization. Another aspect observed was the different optimizations found for each model, even in the same scenario. An additional interesting result was the possibility of the reuse of optimization between scenarios, presenting an interesting approach to reduce time and computational resources.