[LOCKSUIT]

Explain to me how a spider-robot would learn to crawl on the floor.

[Korrelan]

Explain to me how a spider-robot would learn to crawl on the floor.

I know what’s coming but…

Your question implies that once the robot is set toward the task no other human intervention is allowed.

The commonest method is to combine both the power of a neural net and genetic algorithm/ evolution.

The only pre-set required is target value (say speed forward). 

The neural net is used to move the legs guided by a servo map, the servo map takes its operational parameters for a simulated G-Nome.

A generic servo/ action map is generated along with forty nine or so (the more the better) copies with slight genetic differences.  Each map is loaded into the bots neural net and the servo map is run.  A ‘fitness’ measure variable is calculated for each map from how well the map reached the target.  The two maps with the highest ‘fitness’ at reaching the target are then ‘bred’/ combined/ averaged and another forty nine copies are produced with slight genetic variations. Rinse and repeat until the highest fitness measure obtained.  This is not a perfect solution but does a pretty good job.

Note: This method leverage's the strengths of neural nets and genetic algorithms to enable a bot to learn a task with no human interaction. 

It is NOT the same as generating random servo actions on a single model and hoping your changes result in a better fitness measure. This is like putting all the components for an engine in a box and shaking it, hoping they will assemble themselves.

 

Edit: Whilst I find the topics title is both arrogant and patronising; it still stands. So please think carefully about your answer. I would hate for you to fail this simple test.

[Art]

You might consider the terms, Movement Forward, as speed is not necessarily being tested early on in the experiment.

I did find many examples of this and most are by Trial & Error...(perhaps a digital Reward and Punishment / Failure) scenario.

Here are a few to help illustrate to others:


With damaged leg(s)
http://www.popsci.com/robot-learns-walk-damaged-legs

A nice paper on learning to walk
http://groups.csail.mit.edu/robotics-center/public_papers/Tedrake05.pdf

[keghn]

 @art you know where the information is at. Hod Lipson was doing amazing work in this area of robotics and
artificial intelligence. But in 2015, he went in a totally went in different direction? 
 What he talk about, for me is ground zero. A place where a AGI will begin, first time turn on, to begin it ascension.

 

 


Josh Bongard mentioned in the video, that @art presented, has his own online robot school called LudoBot.
  

https://www.reddit.com/r/ludobots/wiki/index#welcome 

http://www.cs.uvm.edu/~jbongard/ 

[LOCKSUIT]

Just as I thought.

You all have a case of Beginner Blindness to how it is actually working, and the correct terms. Kinda like "the plane flies because it has arms, it goes up, and the wind helps". Just imagine how well understood the people latter would become and explain it.

The answer should have been as follows:

1) You place the robot on the floor.
2) Turn it on.
3) It starts generating random actions for all motors, and are saved as part1 of a single scene.
4) Sensory input passes rewards and happens to trigger the accelometer reward.
5) This scene of senses (part2) is linked to the actions tried, and ranks them.
6) Now when sensory input enters again, it searches and selects the highest ranked but similar sense and initiates the linked actions.
7) It moves forth.
8) The actions it just selected were tweaked, this continues improvement. (only on a pc can you run ex. a 100 different copies off its body guessing/tweaking at the same time).

[kei10]

Just as I thought. You all have a case of Beginner Blindness to how it is actually working, and the correct terms.

Beginner Blindness?

1) You place the robot on the floor.
2) Turn it on.
3) It starts generating random actions for all motors, and are saved as part1 of a single scene.
4) Sensory input passes rewards and happens to trigger the accelometer reward.
5) This scene of senses (part2) is linked to the actions tried, and ranks them.
6) Now when sensory input enters again, it searches and selects the highest ranked but similar sense and initiates the linked actions.
7) It moves forth.
8) The actions it just selected were tweaked, this continues improvement. (only on a pc can you run ex. a 100 different copies off its body guessing/tweaking at the same time).

Mind I remind you that the practical mechanism you gave us is somewhat identical to korrelan's post and the others; Teaching a robot walking (either by itself or by supervision).

The only difference is that there are many ways to do it, many algorithms, greater algorithms and mathematics that you couldn't achieve because you claimed that you do not use math and such.

We don't have to stick to your classic sense-rank-action algorithm because there are better alternatives that worked very nicely using various mathematical models that based on machine-learning. Many has done the same using those mathematical models.

Please learn to alter your stance before alleging any unseemly words.

You have the skills to build that baby robot using K'nex. Yet, why don't you create more of that to show us? We have never said we weren't interested to see more of your robots that were constructed with your algorithms. You said you are busy working with AI all the time -- Now I wonder what do you actually work about...

[Freddy]

@LOCKSUIT,

It's called trial and error.

When I was in college back in the 80s we did something similar but used black tape - the idea was the robot would follow the line of tape.

The robot moved it's sensor around randomly, if it found black it went that way, if it found white it didn't. Much like your walls. It didn't really learn anything as such as I don't recall we made it remember anything. But that's just an extra step that we could have done on our BBC micros, 5 1/4 floppies and all, if we'd had the time.

Looking back we could have stored successful travels and then removed the tape once it had learnt the route. Then essentially have it play back the moves.

Yours is a good enough explanation of how it works though in this extended way, I see nothing wrong with it, it makes sense. 

[Freddy]

Here's a whole load of maze learning robots that I have been holding on to. See what you think of all those.

http://cyberneticzoo.com/tag/maze-learner/

[Freddy]

Reading back again I see Art already mentioned that it is trial and error. 

Art, does a Roomba map out a room or is that some collision thing ?

[LOCKSUIT]

I didn't put the turn actions nor wall cue into my Lego robot to be acted on cue, the actions and cue was learned. And the things I skipped as said in the video I have now fully understood.

I think, I cannot make better Lego robot, to do more it needs a more language-able body if you get my hint. I have all this why and mazes explained in my database separately, I carry all this knowledge.

Kei, there is not even one other algorithm that can match this one, because you have to generate actions, and have input search for sense cue match in memory&do linked actions and save senses here since the input looks alike [KNN]). The KNN is part of this greater evil. You couldn't come up with even one. Also, this is functions, math just runs functions.
Go ahead kei, tell me another way it learns to crawl on the floor forth. It will involve actions, generating, senses, rewards, saving, searching & initiating.

LOCKSUITS are for when you need to stop a AI. Watch it doesn't break loose, keep it strapped up on a bed. Even if it just wiggles a bit, it will soon break free. They're also good for when you sleep at night and don't want the dream creators to move your body and sleepwalk, notice how you jiggle in bed, thankfully the drug halts motor signals.

BABE the pig.

[Korrelan]

Excellent posts and vids.

Good ol' 'Motor Babbling' has been a technique for as long as I can remember. Brings back lots of memories lol.

https://www.google.co.uk/webhp?sourceid=chrome-instant&ion=1&espv=2&ie=UTF-8#q=motor%20babbling

 

[Freddy]

I had not heard it called that Korrelan, thanks, I learnt something new .

[Art]

@ Freddy - "does a Roomba map out a room or is that some collision thing ?"

Most Roombas use a few programmed paths to eventually cover an entire room but they rely heavily on external bumpers to let the bot know it has collided with some object and to try another direction. I have heard that their latest model, and there's been a lot of model versions, is finally using laser assistance similar to Neat-O vacuums.

The Neat-O Robot vacuums, on the other hand, make use of an onboard laser to "map out the room" as well as to locate objects and avoid them. Given that everyone's rooms are different with a vast amount of configurations and furniture placement, it also has bump sensors for tight conditions. It also has a square front to allow for close corner cleaning.

[Art]

@ Locksuit - Beginner Blindness? Really?

You should realize (perhaps you actually do) that not every one at this forum or who visits this forum knows or is interested in the actual inner workings or code but rather in the general idea itself. A picture or video says a lot in an illustrative way. I thought some might enjoy them.

I could have described a lot of the inner workings and associated circuits and sensors that enable a four legged bot in the example to craw or to learn to craw but that might only bore any visitors or others not familiar with electronics or the inner workings of a robot and others simply would not care.
After all, this is a discussion among friends, not a TED talk attempting to educate or illustrate techniques to a classroom of scientists.

Yes, I am quite familiar with planes and how they fly, having flown real planes before you were born.

Do be careful of your chosen words lest you have to eat them. 

[LOCKSUIT]

Why would you explain how using such terms? My 8 step explanation is in English and is the only algorithm that could learn to crawl.

Also would you like if I got a baby hamster, put it in a box with food etc, filmed it with a surveillance camera 24/7 to a 2TB HDD, and narrated it?

Also what do you think their spider robot in the video does when it gets to the wall? What do you think about this/think the creator thinks?