Conveyer belts?

[JoshKlint]

How can you make a conveyer belt with Newton?

How can I make it handle curves along a spline?

[Sweenie]

For a straight conveyor a fixed kinematic(static?) body there you just set it's velocity manually would probably work.

Otherwise I think contact manipulation is the way to go, especially if you want curved conveyors.
Not sure which version you are using but the demos BasicFriction/BasicFrictionRamp in either version should be a good starting point for using the contact callbacks.
Inside the contact callback you then modify the contact velocity either with the single direction of the conveyor or based on a spline curve.
The curved part may be tricky though because the velocity would be higher in the outer part of the curve and lower on the inside.
If you set the same velocity and direction of all contact points a box would not turn in the curve I guess, but just get a linear movement like it's floating in water or something like that.

[EDIT]
Actually, might be worth trying the "real" thing.
A series of thin boxes following a spline using custom fixed joints.
Just make sure the joints are stiff enough and the conveyor boxes are not colliding with each other.
And then a body reach the end of the spline just move it to the start.

I think there is a demo there a bunch of boxes follow a spline.

[Julio Jerez]

hi Sweenie.
Long time no see.

I have been working on deep learning.
I found out that Depp learning once you depart form the cooky stock demos,
It is not as easy as they make you believe.

I have seen problem in with you change a parameter from 0.005 to 0.00501 and the whole thing completely collapses.

I am actually developing my one Library. I find tensor flow and py-tourch too combersome.

anyway, I do see lot of potential in use deep neural net as controller.
I have had minor successes, in a series of training seminar that I made myself.

each test has a degree of complexity and are kind of learning tool to learning the tricks.

Julio

[Sweenie]

Hi
I always check the forums now and then but it has been a busy time at work.
I work at a big food manufacturing company and we are migrating our warehouse system. A large project that has been going on for well over a year now.
After a full day of coding and troubleshooting I don't have the energy to do any hobby coding. I just fall into the sofa and watch Netflix
But we are getting close to Go-Live now so looking forward to start coding on my hobby projects again.

We are actually using AI to some degree, not for something cool like teaching a robot to walk but we have an AI that knows how a perfect pizza should look like and sort out the bad ones from a conveyor belt

[Julio Jerez]

interesting.
I always had the impression that you were somehow related to the Software industry.
By this, I mean, working as a develop of software for third party consumption. \
Of course, many people do work in software, but not as the company product, software is a tool.
anyway, yes, we are getting old, and now is time for the new generations to be our replacements.

I do not have the energy that I used to have anymore.
in the past two years, I turn my attention to machine learning, but not for the high level, but from the fundamentals.
I am really appalled at how much misinformation there is out there regarding to deep machine learning. But what else is new.

my interest in deep learning is for using a supplement tool, not the other way around.
I am interesting is using is as controllers for simulations.
so the change I made to the physics engine are to satisfy functionality required for interfacing with the deep learning library.

this is a very slow and tedious process, because I have to do everything be myself. but no one is waiting for me.

My experience with machine leaning so far, is that when it comes to controlling real things, it is not the slam dunk that people think it is.
Reinforcement learning based on deep neutral nets, excels at stuff like large language model, image generation, classification, and stuff like that.
But is is much harder when trying to use for physic. Simulation.
Reason is that need mural net are just big interpolator and extrapolators from data points.
So when the data point are factuals, meaning you get it from an object e source, them the prediction are more reliable.

When the data comes for a simulation, the data point can be very unreliable, so the generate result are just as good as the input data.

so the whole idea is to train the net to generate interpolations that are rewarded high by some objective function and discard the one that are reward low.

people have been doing that unsuccessfully for decades.
-Nearest neighbor interpolation.
-Principal component Analysis.
-Gaussian processes
-Linear regressions
-Lineal control theory

are just some of the few methods. I experiments with few of them, but I have to admit they do not come close to what a deep neural net can archive.

the only problem with it, is that when it comes to physical problem, it is not true that given two different physics state of a configuration, and interpolate state is also another valid state.

so, training robots, is still and unsolved hard problem, albeit it is way simpler to talked using deep neural net that using these other methods.
anyway, that's what I am doing now.

We are actually using AI to some degree, not for something cool like teaching a robot to walk but we have an AI that knows how a perfect pizza should look like and sort out the bad ones from a conveyor belt

exactly

[Sweenie]

So basically you are trying to create some kind of "MacGyver" AI that excels at handling unexpected conditions

I am 100% a software developer but all the software is for the company only.
Anything from printing solutions, deviation management system, product planning system, customer care, workflow systems and so on.

A few years ago we were contacted by a company that wanted to sell us a document & contracts archiving solution.
It was a professional looking system and very expensive as well.
I'm normally not involved in such meetings but for some reason I was invited to that meeting by our financial manager.

So the sales guy did his presentation and then the financial manager turned to me and said, "Can our existing archiving system that you created do all these things?"
"Most of it" I replied.
"Can you add the missing functionality if we wanted?"
"Yes", I replied.
And then the FM told the sales guy that we have no need for their system.

I still don't know why that meeting ever happened because I know the FM already knew that we didn't need to buy that software. I think the FM just wanted to humiliate that sales guy for some reason.
When we left the meeting room the sales guy whispered to me, "You are working for the wrong company"

I know what he meant, I could probably make a lot of money by going commercial and sell these tools instead but that is my big issue.
I love creating all kinds of software, the actual problem solving part of it. But as soon as I'm done I don't want anything to do with it anymore.
You will need to support it, make documentation, update it, market it etc and I find those things extremely tedious and boring.

I recently got diagnosed with ADHD(or ADD since I lack the hyper part)
So that probably explains why I only like to do the fun parts of it and hate the boring parts

Sorry Josh for derailing your thread

[JernejL]

I think the FM just wanted to humiliate that sales guy for some reason.

I'd say this could have been a sales tactic, to get the price down to something reasonable.

I'm glad i see you work for a reasonable company, we regularly buy specialized software that could be developed in-house in a few months of work max.

[Sweenie]

Yeah, I really like this company. I've worked there for over 25 years now.
It's like my second home

I have saved them tons of money over the years by developing so much software in house but I have also warned them of the backside of it.
Don't let me create any business critical software even though I could.
We are a couple of developers but with different areas of expertise.
Even if I try to document and comment as much of the code I can, if anything happens to me they could be in trouble.
It's the classic "Bus factor" problem.

So while my tools makes some processes faster and easier, always make sure you can still do it the "manual way".