Slashdot Mirror
Scientists Create Robots That Can Assemble IKEA Furniture For You (sciencemag.org)
sciencehabit shares a report from Science Magazine: Although artificial intelligence systems may be able to beat humans at board games, we still have the upper hand when it comes to complicated manual tasks. But now, scientists have created robots that can do something even most humans struggle with: assemble an IKEA chair. Putting together a chair requires a combination of complex movements that, in turn, depends on such skills as vision, limb coordination, and the ability to control force. Until now, that was too much to ask of even a sophisticated robot. But researchers have finally broken the dexterity barrier by combining commercially available hardware, including 3D cameras and force sensors, to build two chair-building bots. To construct their IKEA masterpiece, the robots first took pictures to identify each part of the chair. An algorithm planned the motions the robots needed to manipulate the objects without causing any collisions; two robotic arms then performed those actions in concert. Feedback from force sensors also helped: When the robot needed to insert a pin into a hole, for example, it would slide the pin over the surface until it felt a change in force. The robots were able to put together the chair in a little over 20 minutes, which includes the 11 minutes and 21 seconds of planning time and 8 minutes and 55 seconds of actual assembly. The findings have been reported today in Science Robotics.
Ikea furniture is great, but imperfect by design. Tolerances are wide, parts vary, and it takes a wack now and then to get the parts into place. This is intentional since it's far, FAR cheaper to build out of particleboard and holes in that are never going to be totally exact.
What is more impressive is an assembly AI that can cope with that. One that can tighten 50 screws slightly differently because they need to be or tweak two pieces so they slot together as intended. Usual laughs aside, Ikea stuff isn't rocket science to assemble as long as you actually pay attention. Their instructions are usually very specific, but no one looks at the details. I've built tons of it and every time I got stuck or confused on some bit it's because i didn't look at the instructions carefully enough and swapped a part/pin/order. Once you figure out their general ways though you can practically ignore the manuals.
Getting a machine to adapt to a repeatable assembly with moderate variations is more impressive than one would first believe. I'm curious how fast round two assembly went and how fast someone who knows how to use an allen key built it instead of those two women who were more interested in smiling and laughing than knowing how to assemble things.
You can get rich if you own a politician, but you have to be rich to buy one in the first place.
It seems to me an AI with the sophistication of Watson should be able to read the instructions, that a trained modern image recognition system should be able to provide part locations, and that a flexible industrial robot should be able to create the appropriate program (including initially unpacking the parts). So it's just a matter of convergence.
captcha: positron