On-line regression algorithms for learning mechanical models of robots: A survey

Abstract

. On-line regression algorithms for learning mechanical models of robots: a survey. Robotics and Autonomous Systems, Elsevier, 2011, 59 (12) AbstractWith the emergence of more challenging contexts for robotics, the mechanical design of robots is becoming more and more complex. Moreover, their missions often involve unforeseen physical interactions with the environment. To deal with these difficulties, endowing the controllers of the robots with the capability to learn a model of their kinematics and dynamics under changing circumstances is becoming mandatory. This emergent necessity has given rise to a significant amount of research in the Machine Learning community, generating algorithms that address more and more sophisticated on-line modeling questions. In this paper, we provide a survey of the corresponding literature with a focus on the methods rather than on the results. In particular, we provide a unified view of all recent algorithms that outlines their distinctive features and provides a framework for their combination. Finally, we give a prospective account of the evolution of the domain towards more challenging questions.

Autonomous online learning of velocity kinematics on the iCub: A comparative study

“…1 For example, elastic tendons actuating the joints may loosen in time, inducing small errors in the kinematics models which would be generally difficult to detect and automatically compensate for. In such contexts, Machine Learning (ML) techniques supersede classical approaches grounded on accurate exhaustive modeling [2], [3], [4]. However, the application of such techniques to online learning embedded on a real robot performing a visuo-motor task are not so common.…”

Many regression algorithms, one unified model: A review

“…As outlined by Sigaud et al (2011), lwpr is particularly interesting to perform regression when the data lies in a limited domain -because it adds receptive fields only in this domain -in a space with a high dimensionality -because it uses nipals to infer reduced linear models. Therefore, it has often been used to learn mechanical models of robots along trajectories, see Sigaud et al (2011) for a survey.…”

“…Therefore, it has often been used to learn mechanical models of robots along trajectories, see Sigaud et al (2011) for a survey. lwpr has many of the same meta-parameters as rfwr, and thus the same difficulties in tuning these meta-parameters apply.…”

“…A key feature of xcsf comes from the insight that the local linear models and corresponding Gaussian functions do not need to be defined over the full input space, an insight which can give rise to much more compact models, see Sigaud et al (2011) for details.…”

“…As lwpr, xcsf suffers from the necessity to empirically tune several meta-parameters, though some of them can be set more intuitively in xcsf than in lwpr. See Sigaud et al (2011) for further discussion and Droniou et al (2012a) for an empirical comparison.…”