Couple of things.

You are not DIY'ng this.

The tolerances you are looking for are tight.

A $30K Artec Leo which is a pretty nice device won't pick up that .05mm because it's accuracy, under perfect conditions is, 0.1mm.

Accuracy of these devices is dependent on a variety of conditions including the materials reflective properties and how much energy they reflect back, the influence of surface reflectivity.

They run the gamut of Diffuse, think concrete, tends to scatter the energy and less comes back to the sensor.

Specular, shiny, metal, mirror, makes nice tight return energy groupings but can miss the sensor completely.

Reflexive, sends back the most energy, stuff like road signs and road paints that light up at night.

I just don't see a way to build or even buy what you are looking for that will generate the efficiency required to pay for itself.

Machine learning.

He'll blow more than $30k just training the models and we haven't even gotten to hardware yet.

I'm not in the biz - only barely biz-adjacent but I think there's something called "on tool metrology" that actively measures dimensions while work's being done. With some cleverness you could "sneak up" on a surface and work from there.

I think that you're not even in "a house in a decent neighborhood" pricing though. More like "a house with servants quarters". We're doing that in our very expensive corner of the defense industry because, duh, other-people's-money.

I'm doing CT radiography now and there could be some cool ways to use that, but you're again talking hundreds of kilobucks just for the looking, before you ever get to the interpretation or the actual doing. Done right however you get precision into single-digit micrometers. Would be a brilliant way to get exact copies of something.

(national labs are not the real world)

What I've learned of some modern white-light interfereometry tech absolutely broke my mind until I actually understood it. Like "does not follow the laws of physics/imaging as I was taught them" but it's amazing stuff.

Barely even cutting edge anymore. I think there's a Keyence system that's pretty affordable (for a research lab) that can do z-axis into the nanometers, on a countertop, in the air in a normal lab space. No SEMs, vacuum chambers or levitating/magic/unobtanium tables necessary. I'm sure Zeiss and their peers have something similar.

If I am reading this right you are trying to take imperfections out of items that are now-nonstandard. To do this with automatic machines I would think you need to add a bunch of zeros to your number above. There is a reason hand finishing is used and that is you can hire a lot of hands for what it would cost to begin to engineer what you are looking for.
Machines are great for repetitive processes where you want many of the exact same, not so good for one offs with lots of variables. Man is still the go to for that.

Use image recognition instead of a laser. Mark the areas with black sharpie, and the machine massages those spots.

If a 'lego computer' can solve a rubik's cube or sort legos, your device should be able to sand marked spots.

100% NOT a civil engineer.

my eyes glazed over reading the original post. :)

I talked with my friend who has done a lot of programming work using lasers to find things and like many here said it really can't be done easily. Thanks everyone for the interesting ideas and reality check.

If nothing else some recent machinery purchases have made the entire restoration process significantly easier and faster.

Did do a little thought experiment on the math. At $45K, it would break even in one year.

I would think utilizing a GOM or FARO style laser scanner somehow communicating with a CNC machining center is what you are talking about. Laser scanners are extremely accurate, we actually have 2 here and they scan to the micron level. The key would be taking the point cloud that these produce and somehow making a machining center understand this by converting to some sort of surface recognized by the CNC machine.
Is this possible....maybe....but you are looking at a huge investment.
We scan part surfaces all the time here and for tooling repair work, we can download this file into a CAD software and create the surfaces we need to duplicate the shape. So really this technology exists. How to actually use the scanner to teach the CNC machine....thats next level.

If you can do that, it should be relatively easy to tell the CNC to remove .5mm over the entire surface which would work well for my needs. You would have to create an algorithm to define the surface but that is just removing points in the cloud (points that define pits) to create a perfect surface.

Actually not that straight forward.

The reference model and the piece that needs work would need to be 100% identical, no variations.

At that point you might as well scrap the idea of the reference model all together and scan every work piece.

But the time and effort for that would most likely not offset the savings of hand working the piece.

There are plenty of existing robots that can be programmed for finish work.

But they don't really care where they are starting from and are working to a known consistent end product.

It seems your starting point is critical to getting the correct end product and since it's variable it's tricky.

This

Things are getting there very quickly. Ten years ago we could get CT scans of irregular (i.e. real) parts and piece them together into solid models for further work, physics simulations in our case, but it sometimes took many hundreds of nerd-hours to process each one.

Nowadays we can do it in minutes (OK maybe a nerd-hour). I've led projects using that tech to eventually make $10M disappear in less than 1/1000th of a second. Super cool to wrap the whole world around and actually use our supercomputers to deal with reality rather than idealized models.

I might have lost the focus here - as much fun as it's been.

Dude with a buffing wheel? How far short of the goal is that?

;)

(edit to echo Tobra/jrj3rd...)

The piece that needs work is the reference model. They are one in the same. That's what is scanned. Every piece of work will be scanned because each one is unique.

You clearly have not hand worked the piece(s). :) Seriously. That's where ALL of the time saving is: having a "robot" do all initial prep and surfacing.

Robots can't do the finish work. That requires hand work and a trained eye. You have it reversed. Roughing out a surface is what a robot is good for. Creating a perfect final surface is what humans are good for.

In a v1.0, the end product is very simple. Remove .XXmm of material across an entire surface that has been scanned.

The accuracy and quality of the model the robot will work from is only as good as the scan which will without a doubt require manual clean up and that clean up can yield a model that no longer accurately reflects the object that needs rough work.

Trust me, the skill set required to clean that scan is going to cost you more than what it costs to do the rough work by hand.

It's every bit as much an art as the hand finishing plus you have the science of it all to contend with.

If you were working with an organic material like wood or stone it would be easier but you are working with metal which introduces all manner of complications.

There is a considerable amount of man hours that will go into it per piece at a much higher hourly pay rate because of the skill set required.

Remember, we built a company around the science of this and I would not consider what you want to do for more than 30 seconds, at least not with what's available today.

I see what you are saying but in my particular application, the surface (metal and cast Mg) is irrelevant. It is on any surface that will ultimately be modified or sacrificed.

What's the best color/type of surface to be scanned? Matte black? If shiny metal is the worst, matte black is probably the best. That's literally 10 seconds to apply a coat of matte black spray paint to create a perfect scanning surface.

Cleaning a scan will be done programatically, automatically. The best thing about software is you write it once, pay for it once and keep selling it or using it over and over and over.

I'm going back on what I said earlier. In its simplest form, scanning an object, creating a reference surface and then telling a CNC to remove .5mm of material over the scanned surface can't be that difficult.

<iframe width="600" height="350" src="https://www.youtube.com/embed/-6QxUEt5mlA" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>

Great example of over engineering a simple task for no gain.

You'd get the same result in the same or less time if you simply took a hi res pic with scale bars and brought into a vector based editor like Illustrator to generate the file to feed to the CNC.

If it's a one off part it would be quicker to print it out on paper and spray mount that to the gasket material and cut by hand.

If it's a repeatable part where you need hundreds you make a die and stamp it out like a cookie cutter.