Here's much the same job, being done almost 50 years ago, by a robot at the Stanford AI lab.[1] This robot has both vision and force feedback, and uses them to assemble an automotive water pump. It does the coarse alignment visually, and the fine alignment by feel.
You're just grumpy :-). On the plus side the computer that is controlling the robot isn't a DEC-10 in a climate controlled room so there is that.
The correct feeling here though should be compassion, here is a group that has been safely nestled in the arms of Google X and is now being pushed out of the nest like so many projects before it, which currently has one such company, Waymo, that is currently not yet dead. Statistically speaking, it is unlikely they will be able to pupate into a products company before they run out of time.
That said, it is also a truism that the constraints on robotics 50 years ago are not the constraints on robotics today. Re-implementing those ideas which had merit before but lacked a sufficiently robust ecosystem to be practical might in fact be really useful today. One hopes that they have the perspective of the excellent technical reports that SAIL produced to guide their development.
It's hard for me to compare precisely, especially since the Intrinsic videos are sped up, but the one you linked looks very shaky and hesitant, and also the "Ikea challenge" seems like it requires more fine-tuned force-feedback than putting metal pieces together. If I anthropomorphize, the Stanford robot looks like an inept/hungover employee, whereas the Intrinsic robot seems convincing that it's actually accurately aware of what's going on.
Another possible difference -- how much programming time did it take to teach the Stanford robot to assemble the water pump? Sounds like Intrinsic trained the robots to do this with little supervision.
It seems to me that this might represent pretty solid progress, although not exponential/paradigm-shift scale like we've seen in some other industries in that period, and nothing in the Intrinsic videos seemed like it was above par for other automation companies I've seen recently. But since you seem to be in the industry, what's your take on whether they seem to be ahead of the game, or even just realistic, with claims like:
> In one instance, we trained a robot in two hours to complete a USB connection task that would take hundreds of hours to program. In other tests, we orchestrated multiple robot arms to assemble an architectural installation and a simple piece of furniture. None of this is realistic or affordable to automate today — and there are millions of other examples like this in businesses around the world.
It's hard for me to compare precisely, especially since the Intrinsic videos are sped up
Here's a longer version in a larger size, either not sped up or not sped up so much.[1] It's using a simple strategy of approaching the socket at an slightly off angle and then twisting into alignment.
That's a standard strategy. Compare this video of assembling Lego blocks with an industrial robot. Note the little twist moves.[2]
Did the machine learning come up with that, or was it preprogrammed? Did ML re-invent remote center compliance? That would be progress.
Rod Brooks went down this road with Rethink Robotics. They went bust.
You can certainly do what they're showing. It's making a profit on it that's hard.
Thanks for the link. It seems to me that this is rather specialized, and would run into trouble with designs that have a very small base. Also, the number of parts is limited. Imho, this is closer to a printer than to a general model assembly robot.
https://youtu.be/cNxadbrN_aI here are neural networks in the 50s. Something having been tried before doesn't mean it isn't worth ever trying again in a context of much improved technology.
Here's much the same job, being done almost 50 years ago, by a robot at the Stanford AI lab.[1] This robot has both vision and force feedback, and uses them to assemble an automotive water pump. It does the coarse alignment visually, and the fine alignment by feel.
[1] https://archive.org/details/sailfilm_pump