![](bg1.png)
TORQUE CONTROL WITH RECURRENT NEURAL NETWORKS
Guillaume Jouffroy
Artificial Intelligence Laboratory, University Paris 8, France
Keywords:
Joint constraint method, oscillatory recurrent neural network, generalized teacher forcing, feedback, adaptive
systems.
Abstract:
In the robotics field, a lot of attention is given to the complexity of the mechanics and particularly to the
number of degrees of freedom. Also, the oscillatory recurrent neural network architecture is only considered
as a black box, which prevents from carefully studying the interesting features of the network’s dynamics. In
this paper we describe a generalized teacher forcing algorithm, and we build a default oscillatory recurrent
neural network controller for a vehicle of one degree of freedom. We then build a feedback system as a
constraint method for the joint. We show that with the default oscillatory controller the vehicle can however
behave correctly, even in its transient time from standing to moving, and is robust to the oscillatory controller’s
own transient period and its initial conditions. We finally discuss how the default oscillator can be modified,
thus reducing the local feedback adaptation amplitude.
1 INTRODUCTION
Central Pattern Generators (CPG) are biological pe-
riodic oscillatory neural networks responsible for a
wide range of rhythmic functions. They can be made
of endogeneous oscillatory neurons connected to non
oscillatory ones or from the sole interaction between
non oscillatory neurons.
Particularly, they are a great source of inspiration
in the robotics field, for the control of joints in loco-
motion. In general, an oscillatory network controls a
joint angle in both directions, and the phase relation-
ships needed between all joints arise from the cou-
pling between the different networks.
The needed parameters for an artificial Recurrent
Neural Network (RNN) to have a periodic oscillatory
behavior cannot be measured experimentally. This
network is most of the time a relatively simplified
model of its biological counterpart when available.
Only clinical temporal data of joints kinetics and
kinematics can be of use, where however it is difficult
to isolate the real control of a particular joint from the
influence of the others.
In the case of non endogeneous oscillatory neu-
rons, in the litterature, parameters are thus mainly
determined empirically or with genetic algorithms
(Buono and M.Golubitsky, 2001), (Ghigliazza and
P.Holmes, 2004), (Ishiguro et al., 2000), (Kamimura
et al., 2003), (Taga, 1994), (Ijspeert, 2001), compara-
tively to relatively few learning methods (Mori et al.,
2004), (Tsung and Cottrell, 1993), (Weiss, 1997).
Though this is useful with large networks in complex
mechanical models, there are two drawbacks. It is
very difficult in general to isolate the resulting dynam-
ics of the different networks, and to understand their
interaction to each other and with the mechanical sys-
tem dynamics. Also, it is not clear how to modify
such networks in an adaptive context, e.g. in the case
of a permanent constraint change on a joint due to in-
jury.
Based on this considerations, we apply a general-
ized formulation of the so called teacher forcing gra-
dient descent-based learning algorithm, to create an
oscillatory RNN as a torque controller for an inter-
esting vehicle with one single degree of freedom, the
Roller Racer. The RNN is put in a closed loop with
the Roller Racer, such that the vehicle can be freely
controlled, where the RNN can be modified perma-
nently.
The paper is structured as follows. In section 2,
we briefly present the Roller Racer model and we
show how it can be controlled with a torque input. In
section 3 we describe the control system. The sub-
section 3.1 presents the generalized formulation of
the teacher forcing learning algorithm with which we
build the oscillatory RNN as a basic torque controller
109
Jouffroy G. (2008).
TORQUE CONTROL WITH RECURRENT NEURAL NETWORKS.
In Proceedings of the Fifth International Conference on Informatics in Control, Automation and Robotics - RA, pages 109-114
DOI: 10.5220/0001501401090114
Copyright
c
SciTePress