Curiosity question 0

[powermad5000]

Just had a thought. Curious to see the responses to this one. Do you think AI robots could do a better job of grading coins than humans? I ask this because I saw a spot on the news about AI looking at mammograms and having a much higher rate of success determining cancers in the mammograms as well as correctly declaring some of the doctors positives as false positives. So, it got me to thinking if they would be better at grading coins....???

[EagleRJO]

There might be more consistency with grading if they could get it right.  But how would AI evaluate cartwheeling?

[RWB]

"Artificial Intelligence" (AI) is no better than the data to which the system has access and the validity of rules of operation. Quality assurance systems in manufacturing has long been far more accurate, consistent and reliable than human "grading" in most situations. However, TPGs have never made the development investment to adapt existing technology to what they claim to do -- assess the "condition" of coins. This is similar to their rejection of empirical research and scientific method in their "guesses" about unusual specimens.

Frankly, institutionalized ignorance might increase the number of submissions as submitters try to out guess the TPG "experts."

[Coinbuf]

It could be easily done for the modern mint products where subjective factors are less of a factor and it is more about the number and placement of marks.   It would take a large investment of time and funds to produce the software needed, and I think a final human look-see would be wise but it is very doable.   Even classic coinage could have an initial screening by computer for a technical grade and then looked at by human eyes for the market/subjective part of today's grading.   I rather doubt it will happen in my lifetime, but maybe.

Edited August 5, 2023 by Coinbuf

[RWB]

On 8/5/2023 at 11:58 AM, Coinbuf said:

It would take a large investment of time and funds to produce the software needed, and I think a final human look-see would be wise but it is very doable. 

The software has existed for 30+ years, same for optical scanners and compactors. Database would have to be prepared along with the numismatic rules and controls, but large parts of that exist from current workflows and tracking SW. (it's kind of like inkjet printer ink -- where the companies try to scare and scam consumers into buying over priced "Mfg Brands."

[Fenntucky Mike]

How long do you think it would take a robotic arm to: Pick coin from infeed, image on one side, image the edge (full 360), set the coin down, regrip, image the other side, image the edge again (have to image where the grippers were covering the first time), place on outfeed, and process? 20 seconds? How fast are coins graded now?

It'll happen when the ROI is there. $$$

[Fenntucky Mike]

On 8/5/2023 at 3:10 AM, EagleRJO said:

But how would AI evaluate cartwheeling?

A scan of the surface could determine the presence of "flow lines" and determine if they are intact, broken, worn, etc. It would take massive amounts of time to develop but it could be done, either that or have a robotic arm present the coin to a imaging system with lighting and rotate it, still going to take massive amounts of time to develop. In the end, if the robot isn't faster than the current process it doesn't make sense to invest or develop such a system. I'm not opposed to it, but I doubt an automated system will gain much traction unless it's profitable or gains some level of acceptance in the collecting community where higher prices could be obtained on a per coin basis.  

[RWB]

On 8/5/2023 at 4:26 PM, Fenntucky Mike said:

How long do you think it would take a robotic arm to: Pick coin from infeed, image on one side, image the edge (full 360), set the coin down, regrip, image the other side, image the edge again (have to image where the grippers were covering the first time), place on outfeed, and process? 20 seconds? How fast are coins graded now?

1 second, 24/7 M-Sun, and holidays for 3 coins at each second. The SW is much faster; only physical movement slows things. Good workflow automation would eliminate most handling so there are multiple feeding arms and multiple retrieval arms. One conveyor can handle several of the same coin simultaneously (most efficient) or several of different diameters (less efficient). US Mint already has most of this automated as described - and it's largely done inside a clean box. Capsules are filled by automation.

Min 216,000 per 20 hr work day; but using the multiple arm fill & remove you could triple that.

Plus, the machines record all the data and can calculate Q/A statistics, and predict SW or HW failures, etc.

[Fenntucky Mike]

On 8/5/2023 at 8:24 PM, RWB said:

1 second, 24/7 M-Sun, and holidays for 3 coins at each second. The SW is much faster; only physical movement slows things. Good workflow automation would eliminate most handling so there are multiple feeding arms and multiple retrieval arms. One conveyor can handle several of the same coin simultaneously (most efficient) or several of different diameters (less efficient). US Mint already has most of this automated as described - and it's largely done inside a clean box. Capsules are filled by automation.

Min 216,000 per 20 hr work day; but using the multiple arm fill & remove you could triple that.

Plus, the machines record all the data and can calculate Q/A statistics, and predict SW or HW failures, etc.

Not a real expectation (1 sec) in my experience with integrating 7-axis arms in working cells, especially with vision. Not saying the tech isn't out there but that type of throughput is BIG money. NGC just donated  a FANUC LR Mate to a local college, I've worked with those and the FANUC Vision software, there is no-way a robot like that could grade coins in 1sec. A whole different level arm and vision would be need to even get close to that throughput. One sec to pick-in-place, sure, to grade, no way. 

[Fenntucky Mike]

I've thought this through a little more and the time could be cut down if the coins passed under a scanner, camera, or both prior to the robotic cell, scanner could be stationary, camera would have to travel with the conveyor, if a camera was even required. Robotic arm would only have to pick the coin and flip, scanner and camera on the exit captures data from the opposite side. Process would be continuous. Imaging and scanning could possibly be set up for capturing edge data during indexing/travel. With a setup like this the arm is only "flipping" the coin and throughput would be greatly increased, 1 sec could be possible. 

[powermad5000]

Aside from the time needed to grade, does anyone think AI would be better or worse at the 70 point Sheldon scale than humans? This meaning picking up or otherwise noting the lack of fine lines, marks, dings, etc. while not only being able to count the number of them but also seeing the amount or lack of wear to give an accurate grade? If it was also fed data on details coins such as cleaned, whizzed, repaired, etc, could it make a better determination of grade using ANA grading standards, thereby eliminating arguments over the label VS the coin in the slabs?

 

[Fenntucky Mike]

On 8/6/2023 at 3:06 AM, powermad5000 said:

Aside from the time needed to grade, does anyone think AI would be better or worse at the 70 point Sheldon scale than humans?

AI would be more consistent, 100%, it wouldn't eliminate disagreements on assigned grades and attributes by the owners though. 

[RWB]

Any computer system (AI or any other popular term) has to be given parameters for the work. "Grades" would all be within those limits and errors would probably be far lower than human capabilities. This requires empirical-based internal standards for each design type. Luster, and related opinions are not relevant....those are market decisions at time of sale. An automated system is ideal for modern cookie-cutter stuff, and is actually an extension of automated Q/A used by all the world's major mints.

The time to scan, analyze, and report a dollar size coin is measured in milliseconds (or fractions thereof), and one must presume this is done simultaneously for both sides. Also, is there any reason to remove a modern coin from its capsule? Leave it as it came from the Mint. (But must avoid polarized light sources.) This will eliminate handling and possible contamination. Then, use the same capsule inside the certification holder. Saves a lot of time, money, handling and potential damage.

[Jason Abshier]

Keep in mind AI robotics is programmed by Man (human) it can still make errors, although accuracy would probably be better …. However they would need to program this AI robot to grades coins in parameter that’s acceptable for the grade what happens if the robot come across a coin that not quiet in the grade it was programmed to be ? A human would have to look at coin to determine the grade … bag marks , rim nicks , wear and so on will have to be programmed within parameters for AI robot to read it correctly…. therefore !  Grading would still and always be “subjective opinions” with or without AI robotics 

Edited August 6, 2023 by Jason Abshier

[Jason Abshier]

On 8/6/2023 at 2:05 PM, RWB said:

Also, is there any reason to remove a modern coin from its capsule? Leave it as it came from the Mint.

I often asked myself the same thing … why remove it from airtight capsule  … why can’t the grader , grade the coin through a capsule and put the whole thing in slab after grading is done ?? I see no harm in adding the capsule inside a slab ….  when I buy a coin I have to examine it through a scratched up holder to determine if it really worth grade it was given 

Edited August 6, 2023 by Jason Abshier

[RWB]

On 8/6/2023 at 2:07 PM, Jason Abshier said:

what happens if the robot come across a coin that not quiet in the grade it was programmed to be ?

Then it "wouldn't be." The system must not "bend" bin assignment.

Normal system testing should reveal any programing errors.

Look-up tables will also increase image processing speed, freeing clock cycles to handle transforms necessary to fully evaluate the coin or other object.

[Mike Meenderink]

This has already been done at PCGS. Its problematic. The systems get it right 85% of the time on type and details but can't recognize exceptional luster from just luster. Human graders jobs are safe for a while longer. IMO (and many, many others) trying to teach a machine how to recognize beauty (eye appeal) and perfection (strike details/spotless) in a coin is near impossible. The data set required for all die pairs, die states, mints, luster, surface variations, types...etc..etc would be VERY large and time consuming to build. Would you want AI judging fine art? Would you want AI judging the next Miss Universe? Why would we want AI judging coins?  

[powermad5000]

As far as I am aware, luster in and of itself is not an overall factor in assessing a grade. I have a 1923 Peace Dollar that is fantastic in luster and looks like it was just made, but graded MS 63. I also have a 1926 Peace dollar that is rather dull in appearance but is graded MS 65. Just as there are Morgans that grade as MS 67 and be blast white while others can grade as MS 62 PL, or MS 63 DPL. Luster, while desired by collectors, is an eye appeal that affects value at the point of purchase.

The numerical grade is assessed by quality of the strike and wear or lack of wear as well as distracting marks, dings, scratches, etc.

Edited August 6, 2023 by powermad5000

[Coinbuf]

On 8/6/2023 at 1:59 PM, powermad5000 said:

As far as I am aware, luster in and of itself is not an overall factor in assessing a grade. I have a 1923 Peace Dollar that is fantastic in luster and looks like it was just made, but graded MS 63. I also have a 1926 Peace dollar that is rather dull in appearance but is graded MS 65. Just as there are Morgans that grade as MS 67 and be blast white while others can grade as MS 62 PL, or MS 63 DPL. Luster, while desired by collectors, is an eye appeal that affects value at the point of purchase.

The numerical grade is assessed by quality of the strike and wear or lack of wear as well as distracting marks, dings, scratches, etc.

I'm sorry but this is simply not correct, eye appeal (which luster and color are a large part) is most assuredly considered in the grading room under the current market grading concept that the big two employ.    Technical grading which only took into account the quality of strike and number and placement of marks that was the way in the 70's and early 80's under the ANA service is long gone.   Today's TPG grading is more about pricing coins than actually putting an archival grade on the tag.   Without seeing the Peace dollar that you describe it is not possible to comment on it, all I can say is it seems clear that whoever graded it MS63 felt differently about it than you do.

Edited August 6, 2023 by Coinbuf

[RWB]

On 8/6/2023 at 4:20 PM, Mike Meenderink said:

This has already been done at PCGS. Its problematic.

What they tried was not what I outlined. No TPG has ever made a meaningful attempt to automate. A working system, sensibly engineered by educated professionals, would cut staffing while vastly improving quality and consistency. It would also end "gradeflation" unless deliberately altered.

Look for ways to succeed, not models of failure.

Edited August 6, 2023 by RWB

[powermad5000]

@Coinbuf It looks even better in hand than in the pics. I also have many more coins with superb eye appeal but low grades. Many Morgans, Washington quarters, Roosys, and some Franklin halves. If eye appeal was all that was needed to get a high grade, then I would have a lot of MS 65, 66, and 67 coins. I also have a Seated Liberty quarter with absolutely fantastic eye appeal and that only got a 61.

 

Edited August 7, 2023 by powermad5000

[Mike Meenderink]

IMO grading a coin focuses on luster, number of marks and surface characteristics only when qualifying beyond MS 66. A coin must be free from circulation wear to be even considered above AU 58. A coin that is free from circulation wear with heavy luster breaks, weak strike, heavily marked or impaired luster can grade anywhere between MS60 -MS65 (unless cleaned or altered).  Any MS 66 and above grades are given out due to the exceptional condition of surfaces, uninterrupted luster and actual counting of individual marks. An MS 63 grade is appropriate for this coin considering the luster breaks, weak strike and several field contact marks. Nice coin though!

Edited August 8, 2023 by Mike Meenderink