This is an excellent post on how we go from controlled demos to real-world (ie, messy!), high success rate robotics. As many speculated, high quality pre-training and edge-case post-training, drives real-world generalizability.
But holy shit making that theory work in
Introducing ACT-2 Preview
The first robotics model to unify broad generalization with high reliability. A single fine-tuning example can teach Memo a new behavior that generalizes.
Zero shot, real unseen homes, 99% success rate.
