Letting Robots Manipulate Cables(news.mit.edu)
news.mit.edu
Letting Robots Manipulate Cables
https://news.mit.edu/2020/letting-robots-manipulate-cables-0713
20 comments
This is using a camera looking at a perforated gel pad as a touch pressure + force direction sensor.
The variety of uses for "point a camera at it" never ceases to amaze me. And they're often cheap enough and reusable enough to be viable to individuals.
The variety of uses for "point a camera at it" never ceases to amaze me. And they're often cheap enough and reusable enough to be viable to individuals.
Just think of how much stuff humans use "point eye at" for....
While very true, this is kinda using vision to replace other senses :) E.g. touch here, and sound (reading vibration via laser / optics).
Both of which we can also do, but to very fuzzy degrees (thing visibly squishes -> almost definitely under pressure. surface appears to be vibrating -> probably making noise). We don't have much for vision -> taste though (https://en.wikipedia.org/wiki/Astronomical_spectroscopy).
Both of which we can also do, but to very fuzzy degrees (thing visibly squishes -> almost definitely under pressure. surface appears to be vibrating -> probably making noise). We don't have much for vision -> taste though (https://en.wikipedia.org/wiki/Astronomical_spectroscopy).
And also just think at how well humans who lack this form of sensory input still function
If you want to make realistic estimates of the bounds of how well you can do automation tasks without cameras, sure. But not really relevant for figuring out how useful "point a camera at it" can be.
I expected fishing wires thru crawl spaces and stuff.
That's hard for humans.
Related: PlayStation robotic assembly including cable manipulation
https://asia.nikkei.com/Business/Companies/PlayStation-s-sec...
As mentioned by other posters.
https://asia.nikkei.com/Business/Companies/PlayStation-s-sec...
As mentioned by other posters.
I saw something recently about the PlayStation manufacturing plant, and I recall that there were snippets of similar robots doing a lot of this complex cable work. I'd be curious to know if this new work revolutionary or evolutionary in contrast to what Sony's cable automation does right now?
I was wondering where I saw the cable robot last. I didn't know there were already assembly robots that did that kind of stuff until now. I'm assuming the major step here is handling a larger number of cables without re-programming where the Sony factory was only optimized for that one cable in that one step?
Huh this reminds me of the TacTip from the Bristol Robotics Lab. I'm guessing the sensor works in a pretty similar way.
Oh it seems there are some instructions on how to build one yourself: https://softroboticstoolkit.com/tactip
Oh it seems there are some instructions on how to build one yourself: https://softroboticstoolkit.com/tactip
I never even thought about this being a problem until I watched the video. It's amazing the types of things we have to build algorithms to handle that our brain does naturally.
Although brains cannot do it "out of the box". Most kids take years of practice to achieve decent dexterity in even simple manipulation tasks. And learning a new skill like guitar in adulthood makes it very clear that, however sophisticated your capabilities in other areas, you have no innate ability to perform complex physical tasks well without a long grind of practice.
So yes, brains are incredible in what they can eventually achieve, and the fact that the calculation involved becomes hidden from conscious awareness can make it seem "effortless", but when you consider the sum of time and effort spent practicing physical tasks over a typical human lifespan, it's really anything but.
So yes, brains are incredible in what they can eventually achieve, and the fact that the calculation involved becomes hidden from conscious awareness can make it seem "effortless", but when you consider the sum of time and effort spent practicing physical tasks over a typical human lifespan, it's really anything but.
I actually need a robot for replugging USB cables.
usb can be so buggy =(
If it's pre-USB-C, I'd like to know your plan for solving the quantum superposition issue USB connectors. You know, how every USB connector exists in a state that's simultaneously correctly and incorrectly oriented? :P
Now that I think about it, I think the simplest way would be to slightly modify the physical cable and the port so you could physically plug the cable in, but it wouldn't work unless correctly oriented. I had always wondered why they didn't just add another bit to the connector and cable that could allow the controller to determine the cable's orientation, then appropriately shuffle the incoming and outgoing bits so they go to the right places. Isn't that roughly how USB-C solves the problem? I'm pretty ignorant when it comes to most hardware matters. :/
Now that I think about it, I think the simplest way would be to slightly modify the physical cable and the port so you could physically plug the cable in, but it wouldn't work unless correctly oriented. I had always wondered why they didn't just add another bit to the connector and cable that could allow the controller to determine the cable's orientation, then appropriately shuffle the incoming and outgoing bits so they go to the right places. Isn't that roughly how USB-C solves the problem? I'm pretty ignorant when it comes to most hardware matters. :/
The wave function collapses on the third attempt.
This looks cool, but it also looks like the only type of cables they're using are copper wire cables.
I used to work at a hardware company that made and sold fiber optic switches with hundreds of input and output ports. I worked on calibration and test software for these things. Step zero to testing and/or calibrating one of these beasts was to hook it up to a custom-built rig we built that ran the software I worked on. This, of course, involved plugging in hundreds of fiber optic cables.
Besides the difficulties mentioned in the article, fiber optic cables have an additional property that makes them hard for robots to manipulate: inside the plastic jacket, they're literally made of glass. So, you can very easily break a cable, rendering it useless until repaired.
When we detected a likely broken cable, we had a neat method of finding the exact location of the break. We'd take a very bright (10 mW) red laser and shine it down one end of the cable, and see where the bright, red spot appeared on the cable jacket. That spot would be where there was a lot of red light leaking out of the cable, so, that would be where we cut and re-spliced the cable.
The other difficulty is that for a robot to be practical in this use case, it would need to be able to deal with the fact that there are hundreds of cables around. Because they are fragile, and any significant bend in the cable can introduce optical loss, you don't want to just willy-nilly go plugging cables in and not worrying about how they lie.
Oh, and, probably the hardest thing for this type of robot to deal with would be the fact that the cables were not all equivalent. Our rig had an infrared light source on it that powered the whole shebang. Each light source had something like 8 output ports that were very, very bright. I don't remember exactly how much power they put out, but, in spite of not being coherent sources, they were very bright on the IR test cards we used to check and see if things were lit up. Each output port from the light source went into a custom splitter we made that gave us hundreds of light rays that were each just a few percent as bright as the original 8 output ports.
Each of these light rays traveled down a single cable, and we had the capability to turn them on and off individually through some clever optics. That meant we could literally light up one out of the several hundred cables, send that light through the switch, and then measure where we expected those photons to come out.
We needed that capability for several of our tests. This meant that the individual cables were not interchangeable! They were, in fact, labeled, as were the input ports on the switch, so we knew we could plug, say, light source cable 57 into input port 57 on the switch, and then look to see where the light would come out, and how bright it was.
Because they are optical cables, they need to be cleaned of dust before they are plugged in, as well. The whole setup was in a room with HEPA air filtration that approximated a clean room, so there was not much dust, but there was certainly some. If a speck got on one of the connectors, it could really screw up our optical loss measurements.
Because of these difficulties, even if a robot were available that could handle a small number of copper wire cables, I suspect such a device could not deal with fiber optic cable. As a result, our solution was to have interns do the job of connecting and disconnecting the switches from the test/calibration rigs, and avoid disconnecting them for as long as was practical. This is why our rig was both a calibrator and a tester.
If you think about how long this would take to accomplish, consider that cleaning the connector might take ~1 second, and plugging it in to the right might also take ~1 second. Multiply that by 100, and you get 3 minutes per 100 cables. As I mentioned, these beastly machines had several hundred ports. And, besides plugging the light sources into the switch inputs, we had to plug in output cables that lead to an optical power meter on the rig. All in all, around 750 cables would need to be cleaned and plugged in in order to hook the switch up to the rig. 3 minutes / 100 cables * 750 cables is about 21 minutes, which would be about the theoretical minimum amount of time it would take to prep the switch for testing. In practice, it took anywhere between about 30 minutes to an hour. Unplugging the switch from the rig took about half that. So, that's about 45-90 minutes involved in total, just to hook the switch up to the machine we used for testing and calibration!
Unfortunately, it meant that we were employing people to do this for a large portion of their day. We were near a university with a decent math department, and, for some reason, the person who hired interns there really liked to employ math majors. So, we had 2 or 3 college math majors we employed in large part to plug cables in and unplug them when we were done.
I think it is fair to say that particular job will probably not be automated away any time soon.
I used to work at a hardware company that made and sold fiber optic switches with hundreds of input and output ports. I worked on calibration and test software for these things. Step zero to testing and/or calibrating one of these beasts was to hook it up to a custom-built rig we built that ran the software I worked on. This, of course, involved plugging in hundreds of fiber optic cables.
Besides the difficulties mentioned in the article, fiber optic cables have an additional property that makes them hard for robots to manipulate: inside the plastic jacket, they're literally made of glass. So, you can very easily break a cable, rendering it useless until repaired.
When we detected a likely broken cable, we had a neat method of finding the exact location of the break. We'd take a very bright (10 mW) red laser and shine it down one end of the cable, and see where the bright, red spot appeared on the cable jacket. That spot would be where there was a lot of red light leaking out of the cable, so, that would be where we cut and re-spliced the cable.
The other difficulty is that for a robot to be practical in this use case, it would need to be able to deal with the fact that there are hundreds of cables around. Because they are fragile, and any significant bend in the cable can introduce optical loss, you don't want to just willy-nilly go plugging cables in and not worrying about how they lie.
Oh, and, probably the hardest thing for this type of robot to deal with would be the fact that the cables were not all equivalent. Our rig had an infrared light source on it that powered the whole shebang. Each light source had something like 8 output ports that were very, very bright. I don't remember exactly how much power they put out, but, in spite of not being coherent sources, they were very bright on the IR test cards we used to check and see if things were lit up. Each output port from the light source went into a custom splitter we made that gave us hundreds of light rays that were each just a few percent as bright as the original 8 output ports.
Each of these light rays traveled down a single cable, and we had the capability to turn them on and off individually through some clever optics. That meant we could literally light up one out of the several hundred cables, send that light through the switch, and then measure where we expected those photons to come out.
We needed that capability for several of our tests. This meant that the individual cables were not interchangeable! They were, in fact, labeled, as were the input ports on the switch, so we knew we could plug, say, light source cable 57 into input port 57 on the switch, and then look to see where the light would come out, and how bright it was.
Because they are optical cables, they need to be cleaned of dust before they are plugged in, as well. The whole setup was in a room with HEPA air filtration that approximated a clean room, so there was not much dust, but there was certainly some. If a speck got on one of the connectors, it could really screw up our optical loss measurements.
Because of these difficulties, even if a robot were available that could handle a small number of copper wire cables, I suspect such a device could not deal with fiber optic cable. As a result, our solution was to have interns do the job of connecting and disconnecting the switches from the test/calibration rigs, and avoid disconnecting them for as long as was practical. This is why our rig was both a calibrator and a tester.
If you think about how long this would take to accomplish, consider that cleaning the connector might take ~1 second, and plugging it in to the right might also take ~1 second. Multiply that by 100, and you get 3 minutes per 100 cables. As I mentioned, these beastly machines had several hundred ports. And, besides plugging the light sources into the switch inputs, we had to plug in output cables that lead to an optical power meter on the rig. All in all, around 750 cables would need to be cleaned and plugged in in order to hook the switch up to the rig. 3 minutes / 100 cables * 750 cables is about 21 minutes, which would be about the theoretical minimum amount of time it would take to prep the switch for testing. In practice, it took anywhere between about 30 minutes to an hour. Unplugging the switch from the rig took about half that. So, that's about 45-90 minutes involved in total, just to hook the switch up to the machine we used for testing and calibration!
Unfortunately, it meant that we were employing people to do this for a large portion of their day. We were near a university with a decent math department, and, for some reason, the person who hired interns there really liked to employ math majors. So, we had 2 or 3 college math majors we employed in large part to plug cables in and unplug them when we were done.
I think it is fair to say that particular job will probably not be automated away any time soon.
Assuming the fiber optic switches had their ports in the same position, why not just make some tooling to hold the cables in position relative to eachother and turn 600 cables into essentially one 600 pin cable? Or n 600/n pin cables?
Several reasons. Off the top of my head:
* This doesn't get you out of cleaning the connectors.
* You'd need very precise alignment to make this work, and I'm not sure our manufacturing tolerances were up to it.
* We would have needed 20 of whatever setup we used to do this. If you're talking in terms of robotics, that's probably the big deal breaker.
* Cables can still break, and getting them out of whatever holder you put them in could be an issue, not to mention even finding a break buried in 600 cables.
A lot of it really just comes down to "fiber optic cables are fussy and fragile," really.
* This doesn't get you out of cleaning the connectors.
* You'd need very precise alignment to make this work, and I'm not sure our manufacturing tolerances were up to it.
* We would have needed 20 of whatever setup we used to do this. If you're talking in terms of robotics, that's probably the big deal breaker.
* Cables can still break, and getting them out of whatever holder you put them in could be an issue, not to mention even finding a break buried in 600 cables.
A lot of it really just comes down to "fiber optic cables are fussy and fragile," really.
Clean the connectors while they are in their fixture. This opens up the possibility of cleaning several simultaneously.
Making something more precise than a human hand is not particularly difficult. If you don't have the manufacturing capability to do it, there are plenty of job shops out there that can for a reasonable price.
If you can make the connections faster, you can run more in series, so you probably don't need 20 sets. Even if you did, making copies of something is cheap compared to the initial development. External hardware can be set up to service multiple stations.
Many cables moving together in properly choreographed ways are going to be far more robust than an intern fumbling around. If you're making the tooling such that it is difficult to do routine maintenance, then it's a poor tool design.
In general, fussy and fragile things are the best for automation - you're already bending over backwards to deal with it. It's when things are easy for humans, with their amazing ability to adapt to circumstances and immense contextual knowledge, that automation struggles to produce competitive results.
Making something more precise than a human hand is not particularly difficult. If you don't have the manufacturing capability to do it, there are plenty of job shops out there that can for a reasonable price.
If you can make the connections faster, you can run more in series, so you probably don't need 20 sets. Even if you did, making copies of something is cheap compared to the initial development. External hardware can be set up to service multiple stations.
Many cables moving together in properly choreographed ways are going to be far more robust than an intern fumbling around. If you're making the tooling such that it is difficult to do routine maintenance, then it's a poor tool design.
In general, fussy and fragile things are the best for automation - you're already bending over backwards to deal with it. It's when things are easy for humans, with their amazing ability to adapt to circumstances and immense contextual knowledge, that automation struggles to produce competitive results.
I’ve only watched the video here, but I wonder how separable cable manipulation is from the actual insertion task. At some point, there’s a transition between the connector and the flexible cable that, for many kinds of connectors, requires a transition in manipulation strategy. Think about how you have to adjust your grip from the wire to the jack when you plug in a Cat6 cable.
I bring this up because the headphone jack example there isn’t really demonstrative of this difficult modality transition dynamic.
Great work nevertheless!