Noisy Localization of Objects in Reinforcement Learning Framework: An Experimental Case Study on a Pushing Task
Yassine El Manyari, Laurent Dolle, Patrick Le Callet · 2022
In robotics manipulation, for a robot to accomplish tasks that involve manipulating and interacting with objects, it must first be able to recognize and localize the objects in the robot’s workspace. Object detection and tracking is predominantly handled by vision systems. Despite recent advances in the field of computer vision on object localization and recognition, it remains generally difficult to accurately estimate the position of objects in a 3D scene due to the intrinsic noise of image sensors. The purpose of this paper is to investigate and examine how a noisy estimation of objects position impacts Reinforcement Learning (RL) training, the performance of the obtained models and the sim-to-real transfer. The policies are represented by a simple multilayer perceptron (MLP) architecture, and are trained exclusively in simulation. The experimental results show that the amount of noise inherent in object location estimation determines the extent to which the stability and convergence of the learning process is affected. Sim-to-real experiments show that models trained with noise transfer better to the real world.