Newsgroups: comp.ai.neural-nets,sci.engr.mech
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!ix.netcom.com!netcom.com!park
From: park@netcom.com (Bill Park)
Subject: Re: NN for signal processing of a torque-force sensor
Message-ID: <parkCzqMDK.3vJ@netcom.com>
Followup-To: comp.ai.neural-nets,sci.engr.mech
Organization: Netcom Online Communications Services (408-241-9760 login: guest)
References: <3avtn7$56l@info.epfl.ch>
Date: Wed, 23 Nov 1994 20:32:06 GMT
Lines: 51
Xref: glinda.oz.cs.cmu.edu comp.ai.neural-nets:20360 sci.engr.mech:9326

In article <3avtn7$56l@info.epfl.ch> zago@lasensg1.epfl.ch (Lorenzo
     ZAGO) writes:

> I am working on the project of a torque & force sensor for a large
> robotic arm. The sensor shall measure all six force and torque components
> from the signals of a number of strain gauges >= 6 located at convenient
> positions, which however cannot avoid a large measure of coupling.  
> The standard way to process is to evaluate the influence matrix, etc.  
> However I am also studying the possibility of using a neural net
> which will be learning from the test data on the prototype under a wide
> variaty of load cases. It seems to me a possible shortcut with respect to a
> the detail analysis of elastic deformations in a complex object and a 
> way which will have the advantage to consider more explicitely the various 
> sources of signal noise bias, etc.  
> 
> Has anyone an experience to share on such a problem or something similar ?
> 
>      ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>     Lorenzo Zago                          e-mail: zago@elgc.epfl.ch
>    Departement de Genie Civil - Lasen                  
>   Ecole Polytechnique Federale de Lausanne          tel: 021  693 24 96
>  CH-1015 Lausanne                                  fax: 021  693 28 63
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Before you go to a lot of trouble, obtain a copy of some of the early
work on automatic calibration of six degree of freedom force/torque
sensors by students at Stanford's Artificial Intelligence Laboratory
in the late 1970s.

The Stanford people wrote a routine that was part of the arm control
system.  It merely placed the hand in a few different orientations,
then grabbed a bracket on the work table and pushed and twisted in
different directions.  The force/torque sensor was all calibrated and
ready to use in about a minute.  I think all this program needed to
know to start with was the weight of the hand, not even its
dimensions.  Very clever, elegant work.  Very "West Coast," or even,
"California," as we used to say in the robot biz.

In contrast, MIT's robotics people at the Charles S. Draper
Laborartory bought a highly-accurate milling machine as a calibation
jig, built all sorts of clever weights-and-near-frictionless-pulley
mechanisms to apply precise loads, and wrote big offline calibration
programs to process the data collected.  It took them hours or days of
careful "lab work" by an expert to calibrate their sensor, and then
they had to mount it on the arm.  

Very "East Coast," as we used to say. $;o>

Bill Park (MIT, Course VI, 1965)
================================
-- 
Grandpaw Bill's High Technology Consulting & Live Bait, Inc.
