Yes. Most of the work in folding clothes is in garment positioning. If you can train robot arms to uncrumple a shirt from the wash and position it, you're 90% of the way there.
The folding itself is trivial with a folding board (10% of the work)
This imitates a human who does a mediocre job of folding shirts. This shirt will end up wrinkled.
I’d be much more interested to see either:
- a robot trained on experienced clothing retail sales associates (can fold a shirt like this in a single movement), or
- a non-human folding technique optimized for the robot arms (having multiple fingers and more arm+hand mobility is part of what makes the human single-move fold possible).
For context, folding clothes is a traditionally difficult problem in robotics you do as a benchmark for how capable your system is. When I worked for a manufacturer less than a decade ago, we were still deciding against doing it at trade shows because the reliability wasn't high enough yet.
The fact that it's hard to get right is the whole point and why the Twitter post is emphasizing that it's being done with very little work on cheap arms.
They should first have researched how to quickly fold a t-shirt. It can be done in two seconds with little handling and some practise. I guess even by using only two robot arms. See for example https://youtu.be/uz6rjbw0ZA0
IIUC, the controller needs to see what pose the arms are in, run inference on that and compute a next step of movement. If there's motion blur, especially with a simple rolling shutter webcam in the macbook, being too fast breaks the vision? That maybe, plus compute I suppose.
TIL Dynamixel servos can be had for less than $ALL_YOUR_BUDGET!
[1]: https://github.com/jess-moss/koch-v1-1