3 ms·

This is huge. There's been so little progress on manipulation over the years. The 1960s Stanford AI lab videos don't look much worse than current work. Part of

This is huge. There's been so little progress on manipulation over the years. The 1960s Stanford AI lab videos don't look much worse than current work. Part of the trick here seems to be using Gaussian splatting as an intermediate form for vision data. That's a new idea. Gaussian splatting is a brute-force process that doesn't try to generate a full 3D model reconstruction or object recognition. It can be applied to any visual scene where at least two viewpoints are available. This just might get robotic manipulation unstuck. Good target for VC funding.

dinfinity 1y ago
It's not terrible, but I definitely wouldn't call it 'huge'. The learned motions are very oldschool in that the performed actions aren't particularly smooth coordinated movements along different degrees of freedom. The Google robot applying a timing belt presented earlier this year looks much more natural, imho: https://www.youtube.com/watch?v=2AAFiuEP7iE https://www.youtube.com/watch?v=2AAFiuEP7iE (albeit still pretty robotic)