Overview:
Regression analysis, the modeling of functional dependencies between variables given noisy data samples, is a central task in many robot learning problems. Relevant problems in robotics include sensor modeling, manipulation learning, learning value functions for control, learning for planning, and many others. Many regression approaches from statistics and machine learning have been proposed to address robotics-related issues such as online updates, estimation of uncertainty, high dimensionality, non-homogeneous noise & smoothness, and missing features.
In this workshop, we would like to develop a common understanding of the benefits and drawbacks of the different regression approaches and to derive practical guidelines for practitioners to choose the right model for solving a given robot learning problem.
As a second focus, we would like to discuss two major points of criticism of standard practices in robot learning research. Alternative methods are typically evaluated using different metrics and data sets, making it challenging to compare them in a standard way. Additionally, the use of data-driven and machine learning methods may, in fact, not always be the most suitable approach to solving a given problem, especially when considering decades of established work in model-based control that has been demonstrated to be highly effective.
Given this context, we would like to explore what regression problems in robotics can be learnt and to find the subset of the problems that really need to be learnt.
Goal:
Our goal is to draw researchers from the different communities of robotics, control and machine learning into a discussion of the relevant regression problems to be learnt in robotics. The full-day workshop will address topics such as: