A Generic Online Parameter (Re-)calibration Framework Using PPL

Abstract

Parameter calibration is a burdensome yet essential part of robotics development which traditionally was done through manual calibration routines. Over the recent years, many authors have proposed to automatically calibrate parameters directly from the input data, during the operation of the robot. While the majority of methods are discussed within the context of a specific application and benefit from the particularity of the problem, some generic approaches are proposed that are applicable to a wide spectrum of problems. However in practice, they require re-implementation and customization every time applied to a different domain, due to coupling of formulations with the model, e.g. linearization steps, matrix decomposition, closed form solving, etc. In this paper, we exploit the expressiveness of general purpose probabilistic programming languages (PPLs) to build a generic online calibration framework that can estimate the parameters of arbitrary robotic systems during operation. The proposed approach, based on Bayes filter and Monte Carlo methods, only requires model specification and works as a black-box otherwise. Hence, it spans the generality to the implementation aspect of the calibration problem which facilitates a range of new applications, e.g. fast prototyping of arbitrary robots. We show a short PPL program is capable of calibrating kinematic, extrinsic, and noise parameters of a classic SLAM dataset with minimum knowledge about the system and the parameters.