Learning Control

Abstract

This chapter presents an overview of learning approaches for the acquisition of controllers and movement skills in humanoid robots. The term learning control refers to the process of acquiring a control strategy to achieve a task. While the definition is in some cases restrained to trialand-error learning, we present here learning control in a broader perspective, with a focus on the representation of skills to be acquired, and on the different learning strategies that can contribute to the acquisition of robust and adaptive controllers for humanoids.

Incremental Skill Learning of Stable Dynamical Systems

“…SEDSII computes a stabilizing control input from a learned control Lyapunov function (CLF). The CLF is parameterized as CLF = x 2 , where β k , P k , and µ k are learned from demonstrations by solving a constrained optimization problem. SEDSII is very effective in accurately learning complex motions while guaranteeing the convergence towards a unique target, but it is not prone to an incremental implementation.…”

“…Having a fixed set of predefined skills is not sufficient to execute everyday tasks in human populated environments. Programming by Demonstrations (PbD) is a well-established approach to rapidly teach new skills avoiding tedious programming [1], [2]. In the PbD framework, the robot can learn by observing the human behaviour (imitation learning) [3], [4], or an expert user can directly guide the robot towards the task execution (kinesthetic teaching) [5], [6].…”

Automatic Parameterization of Motion and Force Controlled Robot Skills

“…When it comes to small lot sizes and highly customized products, the demand for intuitive programming techniques raises in order to implement new processes quickly. Hereby, the Learning from Demonstration (LfD) [1] technique enables defining a task by nonexperts. However, it has been shown that the robot could suffer from a bad teacher's demonstration performance [2].…”