I like this discussion topic! 
Here are some ideas that colleagues and I are playing with that utilize cheap robots (no cameras or microphones) while still doing very interesting things with AI. 

1. Reinforcement learning: playing simple games like snake, programming itself to navigate mazes

2. Generative algorithms: visual art, poetry, music

3. Character design: teaching a robot how to communicate. Think about everything that a robot like AIBO (robot dog) does to convince you that it's a dog. There are a lot of interesting CS and psychology topics in here

On Sun, Jan 5, 2020, 03:09 Ken Kahn <toontalk@gmail.com> wrote:
I replied to Dave's comment here he posted 3 days but somehow replied only to him personally (I blame jetlag). I think this is an important discussion for this group so I'm copying it and adding a new post.

I wrote

Dave wrote:
>    Also, UBTECH  Alpha 1 S   (robot)
Useless.  Blind and deaf robots have no AI. 
 
So Helen Keller had no intelligence?  
 
A bit more seriously I agree that robots with zero sensors are not good for AI. But even 50 years ago Logo turtles with a touch sensor could be programmed to navigate around obstacles. A simple light sensor could be used to make the turtle act like cockroaches or to follow a light.. One can argue that for children learning about AI and robotics that starting very simple is good and later one can incorporate vision and speech.
 
And then there is this from 1950:
 
SCIENTIFIC AMERICAN MAY 1950
An Imitation of Life
Concerning the author's instructive genus of mechanical tortoises. Although they possess only two sensory organs and two electronic nerve cells, they exhibit "free will"

And Dave replied today:

Hi, Ken.
> But even 50 years ago Logo turtles with a touch sensor could be
> programmed to navigate around obstacles. A simple light sensor could
> be used to make the turtle act like cockroaches or to follow a light..
But no one would call that AI.
> One can argue that for children learning about AI and robotics that
> starting very simple is good
I don't think that's a serious argument.  It's like saying that kids
should learn to communicate by Morse code before being given access to a
phone.
We have 8 year olds doing real-time computer vision and landmark-based
navigation using the Cozmo robot and my Calypso for Cozmo software
framework.  They are getting hands-on experience with real AI right now.
Messing around with light sensors would be a pointless diversion. 

And here's my response:

There are 2 issues here:
  1. Are very simple examples of intelligence AI?
  2. Is it pedagogically wise to start with simpler systems?
Regarding (1) I agree that many examples of AI from the 1950s and 1960s are not called AI anymore. But this strikes me as the AI effect (see this Wikipedia article). Once we understand a problem well enough the program is just a program not an AI program. To me anything that we call intelligent when performed by living things we should be willing to call AI when performed by a computer. Good journals publish many articles about intelligence of bacteria and slime molds.

Regarding (2) there are many arguments for children building very simple AI systems:
  1. It is good science. Biologists know that they can learn about biological systems by studying E. coli or C. elegans or Drosophila that would be too difficult if studying mammals. Physics is taught at first in idealised conditions - e.g. no friction.
  2. Many argue that students should encounter as few black boxes as possible when doing computing, robotics, or AI. Image or speech recognition are black boxes except to a very small number of experts.
  3. Almost everything I've seen from the Arduino, Raspberry Pi, and micro Bit communities begins with one or two very simple sensors. Children can now for example build Braitenberg vehicles
  4. Simple systems are much more affordable as many schools in the world sadly cannot afford Cozmos or like
Note I am NOT arguing that students should ONLY be exposed to very simple  AI  systems. Only that simple systems should be part of the mix. Work like Dave's Calypso for Cozmo and my own work on making many high-level (regrettably black box) AI blocks in Snap! are also important (and perhaps should get the majority of the attention).

I hope our disagreement isn't about definitions of AI. I know some people these days think machine learning is AI. But I like this Wikipedia definition: "...any device that perceives its environment and takes actions that maximize its chance of successfully achieving its goals..."

Best,

-ken

On Sun, 5 Jan 2020 at 14:06, Dave Touretzky <dst@cs.cmu.edu> wrote:
Hi, Ken.

> But even 50 years ago Logo turtles with a touch sensor could be
> programmed to navigate around obstacles. A simple light sensor could
> be used to make the turtle act like cockroaches or to follow a light..

But no one would call that AI.

> One can argue that for children learning about AI and robotics that
> starting very simple is good

I don't think that's a serious argument.  It's like saying that kids
should learn to communicate by Morse code before being given access to a
phone.

We have 8 year olds doing real-time computer vision and landmark-based
navigation using the Cozmo robot and my Calypso for Cozmo software
framework.  They are getting hands-on experience with real AI right now.
Messing around with light sensors would be a pointless diversion.

-- Dave

To unsubscribe from this list please go to http://www.simplelists.com/confirm.php?u=kRqbboLBf1PlKZs7cr2t6Zdg2oDMStj1