Robot Perception

[Auckland Neil]

Hi

Is anyone else interested in the area of robot perception? By this I mean the area of AI concerned with the recognition of incoming data and the correlation with pre-existing expectations about the world. or something like that...

There is a lot of new thinking in neurobiology which should find application here, but i hven't found a good forum.

Is there anybody out there?

Neil Hawkes

[HarryBotter]

Hi Neil,

I'm generally interested in that but beyond doing a bit of reading and trying to get into vision I haven't done very much.

The amount of time required to get up to speed has scared me away a few times but I seem keep coming back.

Can you give me more detail on what you mean exactly when you mention new thinking in neurobiology?

Cheers, Harry

[Auckland Neil]

Hi Harry

OK - brace yourself (^_^).

There is, of course, a lot going on in neurobiology (most of which I know nothing about) but I met a guy called Dr Walter J Freeman a few years ago at an AI seminar where he was giving a paper outlining his work in intentional systems arising from non-linear neurodynamics.

It is very hard stuff to preci - I could say it is a bit like my 'Positonics' stuff (see the attachment) but it arises from an explanation of the mind as arising from the self-organisation of patterns of chaotic 'firing' behaviour in huge populations of neurons.

He (Freeman) wrote a book about it called 'How our Brains Make up their Minds' which is very good, but not an easy read. Anyway, it set me thinking....

There have been some interesting attempts to model such non-linear population dynamics computationally - e.g. Dr Derek Shawn Harter's doctoral dissertation ( I have a copy if you are interested - but it is an 11MB pdf - so I'll not try to post it here! Anyhow, this stuff is super-cool, but way too computationally intensive to be useful for robotics - unless your robot has a supercomputer for a controller!

So - the 'Positonics' stuff came out of my musing as to how similar non-linear dynamics could be engineered (instead of being 'emergent' from a chaotic system) - giving up the self-organising behaviour but keeping the non-linear population dynamics as a way of encoding expectations in a way that could be amenable to modification by incoming sense-data while using hugely-less horsepower than is needed to model the biological systems, so as to give a system that could be useful in 'robotic perception'.

OK - relax, I've finished - if any of this is of interest, see the attachment - and if it isn't, do you know of anyone else thinking about this sort of stuff?

My trouble with all this is that I have had to make up my own language to describe and manipulate the concepts (which sounds pretty weird, I know) - so if there is anyone else thinking similar stuff, they won't have called it 'Positonics' - so how can I Google it?

So - any advice about a likely forum would be very welcome.

And sorry this is sooooo long.

Regards

Neil

[HarryBotter]

Thanks for your response Neil.

Any searching I did in the area of chaotic dynamics usually brought me back to some paper by Freeman. Seems like he's either the leader in the field or the only one engaged.

I doubt that you'll find anyone on this forum active in this field, at least I haven't seen anything close to this here.

I've briefly looked at your attachment and quite frankly don't understand where you're headed with it. I've downloaded it and will take a closer look when I have time.

Cheers, Harry

[Auckland Neil]

Thanks for that.

I know what you mean - it is all very 'left field' but I thought it was worth a try.

As to where this might lead - I was thinking of rigging up a trial-system using a web-cam and an FPGA on the back of a small controller - the Botball XBC2 on the back of a GBA looked a good fit. The XPORT that the XBC piggy-backs on has I/O ports and FPGA development facilities built-in, so I could take the vision data from the camera, process the low-level assoton demography asynchronously in the FPGA and then pass them into the GBA memory for further processing.

If I can get to the point where it can recognise or perceive aspects of it's situation - e.g. walls of a maze and suchlike - then I could strap the lot onto a Robosapien and hack the robot controls to get something quite autonomous.

I am a systems engineer (C.Eng. MIET) so I can do this stuff, maybe when I get the time.......

Cheers

[alain]

Here is an interesting site you might find some answers to your subject :
http://www.geneticprogramming.com/

[D0n]

Can I ask a question?
the seti program uses clusters of computers over the internet... my macs have xgrid...

why do you feel you need to have your super computer on board your robot?

...man, if you could hammer out software that thinks on that level, using many computers over the internet, and have it log into a robot, wirelessly, to control it.....

any number of robots, anywhere, could be used by one thinking program....one mind as it were....

Ever see the movie "Fallen" with Denzell Washington, and John Goodman?

[Auckland Neil]

The idea of using the web for positonic processing is interesting.

One of the aspects of the basic idea is that the information passing from one positon to another (e.g a one-bit inheritance potential within an FPGA) is extremely minimised - this being necessary for any form of massively parrallel processing, or the overhead involved in communication gets unmanageable.

Such minimisation wouldn't be possible on the web, but that is less of a problem if you do the low-level stuff locally and the high-level stuff over the network.

So yes, I guess so.

Many interesting sci-fi stories exist about emergent intelligence from telecomms networks gone awry.............

[TheDuck]

Neil!

I have been interested in something like this for some time. Have you ever read the book Evolutionary Robotics by Nolfi and Floreano (http://www.amazon.com/Evolutionary-Robotics-Intelligence-Self-Organizing-Intelligent/dp/0262640562/ref=pd_bbs_sr_1?ie=UTF8&s=books&qid=1210688607&sr=8-1)? Also, did you read about the "Darwin Chip" in Discover magazine some time ago (http://discovermagazine.com/1998/jun/evolvingaconscio1453)?

I'm also intrigued with the concept of massive network parallelism noted above. Imagine using a multi-node Amazon EC2 cloud to run a robot!

A project on my hitlist is to use genetic programming concepts in a swarm. So imagine 100 robots incrementally trained to clean up a park, search for "different" ocean life, explore a planet, mow a lawn, etc. Also, imagine they are not statically programmed. They constantly interact with each other sharing their "genetic code" and evolve with their environment. They would be learning from their experiences and others and refining their activity based on these experiences. Theoretically this should result in faster and more robust adaptation.

Now, I don't know if this fits into the robot perception group as you are thinking about it in the original post. I think perception in robots will have to include their inputs, processing and outputs and not just their sensors. A robot with a BASIC Stamp and a bumper switch has a different perception from one with a multi-core Propeller and an array of sensors.

But before I go too far along I'll read the attached document and get a better idea of the subject matter.

-Doug

[shimniok]

Interesting... downloaded, skimmed... trying to wrap my head around it... :D

Michael