Autonomous robots and ai ?

AI-controlled robots are extremely complex things. The robots in the hobbyist sector are very simple devices that can react to relatively simple conditions.

I was watching a TV program about a Robotics Research Center here in Hamburg. The professor leading the team said this "We have robots that when connected to a computer can play chess and move the pieces around the board and sometimes win. What we don't have are robots that can go to a shelf, pick up the board and box of chess pieces, unfold and place the board on a table and set pieces up correctly ready to start the game"

Which I believe is a good representation of how far AI/Robotics has come. This is definitely not very far, IMHO opinion.

Furthermore, BBC World featured some robots in Singapore. They were cleaning the floors of a museum there. If someone got in their way, that person would be asked politely to move. These robots could also play tunes and recite poetry.

Your post made me consider the following. If you left a piece of luggage on the floor would the robot also ask it to move out of the way or recognise it as an inanimate object and just move around the obstacle? I guess that would require AI or at least an infrared sensor to tell the robot it was not dealing with a human being because it was not giving off a heat signature.

The border between robotics and simple automation are not very clear. In my professional life, I have come across "so-called" robots that can pick up a piece of metal, hold it in the right position and tack or filet weld another piece to the first. These were the type of robotic arms, such as used also in car spray painting, that are taught in situ how they are to move the spray or welding head. After that, they just follow their instructions 24/7 forever.

AI or just automation?

We say something is behaving in an intelligent way if we can perceive its actions are determined by some goal state. If we observe that a robotic arm will always pick up the red block regardless of where we place it on the table we will credit it with more intelligence than a simple pick and place robotic arm that goes through a fixed sequence of actions regardless of the outcome. Also we credit it with more intelligence if it keeps changing its behaviour (learning) until it finds a set of actions that will take it from the current state to the goal state.

Posted by: @casey

We say something is behaving in an intelligent way if we can perceive its actions are determined by some goal state. If we observe that a robotic arm will always pick up the red block regardless of where we place it on the table we will credit it with more intelligence than a simple pick and place robotic arm that goes through a fixed sequence of actions regardless of the outcome. Also we credit it with more intelligence if it keeps changing its behaviour (learning) until it finds a set of actions that will take it from the current state to the goal state.

 

This method is kind of like subsumption robotics. I love it and it gives you a way to order things to be done.  I used it in my GPS robot years back to keep it on task.  David Anderson has a great video on this that I highly recommend. 

I've been studying A.I. since the 1960's.   Back then A.I. as we know it today didn't even exist.  What did exist at that time was what is known as "Expert Systems".   In fact, at that time that's exactly what they were calling "A.I.".

I actually took a college course on A.I. in the 70's and it was indeed a course exclusively on "Expert Systems".

What is an Expert System?

An Expert System is a computer program that contains a knowledge base of information that can be queried and lead the inquirer to a solution to a problem.   After taking that course I actually wrote quite a few Expert Systems.  One was to troubleshoot TVs.   I did this while working at a TV repair shop.  I programmed it in BASIC using a TRS-80 computer.   It works so well that the other technicians were actually using it to help them troubleshoot TVs.   They were amazed at how well it worked.

As time passed the term A.I. grew to mean more than just Expert Systems.   In fact, I never liked the idea  of calling an Expert System A.I. in the first place.  I was actually disappointed when I took the A.I. course in college and discovered that is was just a course on E.S.   I was hoping for more.

Today, the term A.I. refers to software/hardware that can actually figure out how to do new things that weren't originally programmed into it.   This including being able learn how to play a game better with every game played.  Like in Chess or Go.

Of course, when we get into having a physical robot interact with the real world A.I. becomes an ability to be able to navigate, recognize, and manipulate physical objects.  This is what  most of A.I. today is focused on.

There is, of course, the sector of A.I. that is focused on cognitive communication. In other words, having the robot understand speech and respond with a meaningful reply.  I'm currently working in this area using my own brand of Semantic A.I.   I didn't come up with the idea of Semantic A.I.   It's been around for some 40 years.  There are many different approaches to Semantic A.I.  I do believe that my approach is unique and I'm hoping to share what I'm doing with this in instructional videos over the winter.  Assuming I get my wood boiler repaired in time. ?  I'm not going to feel up to making videos if I'm freezing to death.

I'm also working on designing my own neural networks circuits using hardware instead of software.  I personally believe this is the key to making truly efficient neural networks.  I've had this idea for decades but was unable to implement it due to a lack of technology available.  This is rapidly changing today.  I now see where I might be able to create some crude analog neural networks using SMD op-amps, FPGAs and having these all mounted on small circuit boards that I can design and have made at very reasonable prices.  This has been totally out of the question before.

So anyway, I'm sure this didn't answer the question of the OP.  But this is definitely the field I am interested in.  As I say, I'm working on both Semantic A.I. which I will be calling R.I. (Real Intelligence), and I'll have specific detailed definitions of the difference between A.I. and R.I. , and on trying to design physical analog neural networks on PCBs.   So we'll see how all this goes.

Right now I need to go back outside and work on real life problems.

Posted by: @duce-robot

Shouldn't an autonomous robot do more than just learn its environment shouldn't it be able to perform a routine of more than one thing such as start up pick up one item from the floor open window blind turn on TV press button on coffee maker stop .I'm wondering how hard it would be to do this ? .....since I'm new to this concept 

The examples you gave would require the robot had arms and hands of some kind to grasp items, pull blind cords, press buttons, pull or push a lever and so on...

The hand would have to be guided to the right place and orientation relative to the object to be manipulated.

So work on one problem at a time and think how it might be done with the hope of discovering what general things are required for similar tasks.

Let us say you want the robot to press a button. First it must locate the button relative to itself. What sensors would you need for this? If you have video input you might use that as a sensor. It will allow you to input an image. An image is a 2d array of coloured dots. Inside the computer they are represented as an array of numbers that represent the light intensity and color of each dot. So the question becomes how to determine which dots make up the image of the button to press? Let us say you have a button that is red. Find the pixels with that color and you have your button's position relative to the camera.

So your first step is to decide on what language you want to use to program the computer. There are tutorials that explain how to find groups of coloured pixels.

@casy I enjoyed that video on Blob Detection, an awesome presenter for sure and a useful presentation.  I probably need to the sort of presentation pace that @DroneBots brings to the equation to save multiple rewinds though ?  

The series I think is a good introduction to visual processing algorithms even if you use another language. I first played around with visual processing 10 years ago using another language but no one else using the language was interested and I drifted onto other topics. Back then there was no kinect or lidar only webcams.

What language you use doesn't matter providing it is fast enough to achieve the result you want. If you are using the Processing language then you would want to start at the beginning. I quite like the look of the Processing Language and I see it can be installed on the Raspberry Pi. And I see it supports the gpio pins right out of the box.

https://www.raspberrypi.org/blog/now-available-for-download-processing/

Mmm, processing does look interesting, but I'm getting into overload territory. ?   I will do some blob recognition stuff in C or Python for now.  I will see how @DroneBots DB1 progresses and follow my leader into the areas he boldly goes.   I looked at a video on introducing neural networks I mention in some other post, and whilst it was a good explanation the maths did give my neurones a bit of a workout. (which is to say I did not have the foggiest in that aspect ? )

@casey

Awesome I see what you mean

There has not been much mention of the learning process in this thread. That is where I sincerely believe the difference is between AI and automation.

Before the days of giving small children iPad like tablets or parking them in front of large flatscreens showing TellyTubbies to improve their cognitive and social development (complete bullshit), those of us old enough to remember, toddlers were given a game with various shaped wooden objects to place in holes.

Certainly, if any hobbyist, could, in their basement workshop build a robot to solve this very simple puzzle by AI learning, i.e. with an ever decreasing failure rate, they could certainly also give themselves the title of "Robotics Engineer".

I know it would really impress me!

Posted by: @pugwash

There has not been much mention of the learning process in this thread. That is where I sincerely believe the difference is between AI and automation.

I have also made distinctions between labels.   And recently I've actually come up with some definitions.

As you say we already have several categories

1. Automation
2. Expert Systems
3. Artificial Intelligence
4.  and I'm prepared to add a fourth category of R.I.  Or Real Intelligence.

But let's not just talk in terms of labels, but rather in terms of definitions:

Here are some definitions that I would offer up.

Automation - any machine that can perform a task automatically.

A washing machine is a good example.  It fills itself up with water, washes the clothes, drains itself, and spins the clothes dry.  I've even seen washing machines that are capable of actually drying the close entirely after having washed them.     These machines have absolutely no intelligence.  In fact, at one time they were built using entirely mechanical times and cogs.  Not even a microcontroller required.   So automation, while it used to be impressive, does not equate to intelligence of any kind, save for the intelligence of the human who designed it.

Expert System - A system that can regurgitate intelligent information based on queries. 

Again, these systems are not intelligent.  They are simply programs designed to ask questions and provide "expert information" (i.e. information programmed into them by experts).  They can also be equipped with robotic arms, etc. and actually perform tasks.   None the less, they are not even "artificial intelligence".   All they amount to are sophisticated "Expert Systems".   Although they can appear to be quite intelligent to those who do not know how they are actually programmed.

Artificial Intelligence - A system that can solve unexpected problems that were not programmed into it by an expert.

Actually this is a very fuzzy definition because one could argue that if the programmer actually programmed the system to deal with "unexpected" problems then dealing with unexpected problems were actually programmed into the system by the expert.  I think this is the case in the example of self-driving cars.  The programmers are actually programming into the cars a wide variety of potential reactions to problems that might come up.   Exactly how "intelligent" a self-driving car might be is hard to say.  In truth, a self-driving car is probably still an example of a very sophisticated "Expert System".

So where do we cross the line into my following definition of R.I.

Real Intelligence - The ability to actually understand the context of a situation and what the system is trying to accomplish.

Please keep in mind that this is entirely my own definition.  But one that I have been thinking about for quite some time.  Recently I've become excited about the possibility of potentially crossing into the realm of creating a machine that has"Real Intelligence".   Note: This is not to be confused with consciousness, or even sentience.  It's simply a matter of the  system understanding the context of a situation.  That doesn't mean that there needs to be anyone "at home" within the system.

I've come to this realization recently while working with Semantic A.I. which I am now prepared to call Semantic R.I. (at least if I can get it to do what I think it should be capable of doing).

It's based on semantics (i.e. the definitions of words).  The idea is to not only program the robot with the ability to speak and  understand speech, but to also be able to process the meanings of the words into a coherent context that generates a meaningful goal based on the definitions of those words.

You mentioned the following:

Posted by: @pugwash

Before the days of giving small children iPad like tablets or parking them in front of large flatscreens showing TellyTubbies to improve their cognitive and social development (complete bullshit), those of us old enough to remember, toddlers were given a game with various shaped wooden objects to place in holes.

Certainly, if any hobbyist, could, in their basement workshop build a robot to solve this very simple puzzle by AI learning, i.e. with an ever decreasing failure rate, they could certainly also give themselves the title of "Robotics Engineer".

I know it would really impress me!

This is what the Semantic approach to A.I. is all about.   It about having the system or robot actually "understand" what it's trying to do.   As I say, there is no need for an actual consciousness or sentience to be able to do this.   All that is required is that the system has a semantic context built up around the definitions and meanings of works.   In other words, "Put a round peg in a round hole" actually means something to the system.   And based on the meanings of those words the robot can then figure out how to accomplish the task.

~~~~~~

Having said all of the above, I should point out that it's easy to make up these definitions and distinctions,  Putting them into practice in an actual robot is a whole other story.   I have a lot of ideas on how this can be done.  Unfortunately, working along on a limited budget prevents me from making much progress on building an actual robot.

The other thing too, is that the current Semantic A.I. that I'm working on works very much like human R.I.  In other words, my robot will need to learn from scratch just like a human baby does.   And it may very well take as long to teach it.  Because keep in mind that even humans need to be taught and learn from examples of others etc.  If we took a human baby and could somehow let it develop to adulthood without ever having been taught anything or having interacted with other humans, then it too would be quite ignorant even though it has a brain more than capable of being educated to a high degree.

So even a robot needs to be taught as it grows.   Or at least this is true with the current system I'm attempting to develop here.   I wish I had a large grant and a bunch of graduate students working with me.  I'd have them building an R.I. robot in no time.  And I'm sure that even though I would be instructing them on how to proceed they would be adding their own ideas to accelerate the progress far faster than I could ever hope to do it alone.

Sorry for the long ramble here.  I just finished up working on some perceptron graphing programs and thought I'd check in before going to bed. Goodnight.

Posted by: @pugwash

There has not been much mention of the learning process in this thread.

Learning is an interesting subject I could talk about ad nauseam but my real interest here is not what I might do in theory rather it is about what I can do in practice.

Now if I could afford this robot maybe I could do what you suggest 🙂

Posted by: @casey

is not what I might do in theory rather it is about what I can do in practice.

That sums it up nicely I think.  As @DroneBots was talking about an AI layer for the DB1 robot I followed the discussions on AI to see if I should be gearing myself up to get into a new area to succeed with my robot.   I did find the concept of neural networks and the theory interesting but doubt if it is very relevant for me.  I do see what is often termed as AI is really just an expert system, and even if the system stores data as part of a 'learning process' for reuse with its algorithms this is really only a marketers AI.    And a compressive expert system with learning algorithms also seems too much to contemplate for what a hobbyist could or needs to achieve. (not a driverless car then ? )