A neural framework for adaptive robot control

Abstract

This paper investigates how dynamics in recurrent neural networks can be used to solve some specific mobile robot problems such as motion control and behavior generation. We have designed an adaptive motion control approach based on a novel recurrent neural network, called Echo state networks. The advantage is that no knowledge about the dynamic model is required, and no synaptic weight changing is needed in presence of time varying parameters in the robot. To generate the robot behavior over time, we adopted a biologically inspired approach called neural fields. Due to its dynamical properties, a neural field produces only one localized peak that indicates the optimum movement direction, which navigates a mobile robot to its goal in an unknown environment without any collisions with static or moving obstacles.

Active Learning in Robotics: A Review of Control Principles

“…For instance, in aerospace applications data-driven techniques will have excellent data available for nominal conditions but often no data available for specific off-nominal conditions, suggesting the need for active learning outside of the nominal regime [112]. Since the literature on learning dynamics is very diverse, providing a comprehensive survey would require its own review [113][114][115][116][117]. Instead, here we review a few particular representations of dynamics that are of particular interest to the field of robotics.…”

Artificial Development by Reinforcement Learning Can Benefit From Multiple Motivations

“…The MORL idea transforms the original problem of learning one behavior that is useful in all circumstances into a problem of designing an appropriate architecture for learning and decision making that combines several (probably hierarchically organized) instances or stages of classical RL and possibly other methods of learning or decision making (Oubbati and Palm, 2010 ).…”

Dynamical learning of dynamics

“…on learning with delayed, often unspecific reward [10][11][12]. Another direction of research is on learning of supervised learning: the considered systems typically learn dynamically to predict time series for the current time step given the preceding step's desired output [13][14][15][16][17][18][19][20][21][22] or to track a desired time-varying state variable [23][24][25][26]. The studies assume that a teacher is present during testing to avoid unlearning.…”

“…Approaches to supervised dynamical learning in the literature consider the one-step prediction of time series [13,14] and the approximation of input-output maps [15-17, 19, 21, 22], where the correct previous output is fed in. Other networks could adapt their dynamics to provide negative feedback for control [24][25][26]44], a pretrained oscillation [45] or periodic sequences of discrete states [46]. Learning of supervised learning has also been used to identify the parameters of a dynamical system [47] or perform optimization [48,49].…”