Truck Backer-Upper

Has anyone studied and/or implemented a network training system such as described in "The Truck Backer-Upper: An Example of Self-Learning in Neural
Networks" by Nguyen and Widrow? Here is a link to their article:
http://www.stanford.edu/class/ee373b/truckbackerupper.pdf
They use an evolutionary approach to train a neural network to provide steering control for an 18-wheeler backing from an arbitrary location to a loading dock.
This type of problem is interesting because it must emulate and evaluate the steering control effect on the truck with it positioned in different locations and orientations. The same approach should be useful for a wide variety of applications such as training a neural network as an autopilot.
After reading their article twice, I thought I understood how to do it. But when I began to think seriously about implementing it, the fog rolled in.
This is quite a different problem than training a network on a fixed set of data. It's an interesting hybrid between evolutionary programming and neural networks: The end product is a neural network, but the training involves evolutionary methods.
If anyone is interested in this type of problem, I would like to discuss it.
--
Phil Sherrod
(PhilSherrod 'at' comcast.net)
  Click to see the full signature.
Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload
On Wed, 01 Apr 2009 00:16:44 +0000, Phil Sherrod wrote:

All I can say is that if you were going to implement this for real it doesn't sound a project that's best done with neural networks, at least not once you get to the point of having a realistic trajectory that you want the trailer to follow.
_Choosing_ a successful trajectory may need it, but even that, I suspect, could be better done by more traditional means.
--
http://www.wescottdesign.com

Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload

The problem is that the control system must guide the truck from an arbitrary location and orientation in the parking lot outside the warehouse. So it is not clear how you determine the "realistic trajectory" to get to the loading dock.

What approach would you use?
--
Phil Sherrod
(PhilSherrod 'at' comcast.net)
  Click to see the full signature.
Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload
Phil Sherrod wrote:

You start with the trailer at the dock and move it to the arbitrary position, a relatively easy calculation. Then, since you have perfect memory, you run the tape backward.

Jerry
--
Engineering is the art of making what you want from things you can get.

  Click to see the full signature.
Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload
On Wed, 01 Apr 2009 00:59:16 +0000, Phil Sherrod wrote:

I don't know. There are several candidate approaches that I can think of, starting with a simple linear system that gets the trailer position and direction right in the end then following that trajectory, and going out from there.
I suspect that the best answer would involve some sort of search of a space of several candidate trajectories, but I honestly don't know off the top of my head how I'd do it -- which is why I'm not going to rule out neural nets as being the best (or at least a quite viable) solution.
--
http://www.wescottdesign.com

Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload

I expect Bernard Widrow just might have thought of that one before moving on to a neural net! ;o)

Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload

IIRC it was implemented with adaptive filtering methods, I have seen Prof Widrow's demonstration and it is very impressive. You could at a stretch claim the adaptive signal processing was a sort of neural net, I suppose.
Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload
wrote:

"At a stretch claim... a sort of neural net"?
From their description: "The first stage involves the training of a neural network to be an emulator of the truck and trailer kinematics. The second stage involves the training of a neural-network controller to control the emulator."
The emulator consists of a neural network with 45 hidden units. The controller consists of a neural network with 25 hidden units.
--
Phil Sherrod
(PhilSherrod 'at' comcast.net)
  Click to see the full signature.
Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload
wrote:

I did say IIRC, obviously a case of IIRI ;o)
I saw the demo many years ago, may have been a different system, may just be my faulty memory, but I do remember being surprised about what could be achieved by relatively standard techniques. The important part of my message is that it is pretty impressive and well worth seeing!
Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload
On Wed, 01 Apr 2009 13:40:04 +0000, Phil Sherrod wrote:

If they're simulating something as simple and as deterministic as the truck's kinematics with a neural net, then they've just proved that they have the mindset that All Problems Should Be Solved With The Magic of Neural Nets.
Anything else they do just follows.
--
http://www.wescottdesign.com

Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload

Polytechforum.com is a website by engineers for engineers. It is not affiliated with any of manufacturers or vendors discussed here. All logos and trade names are the property of their respective owners.