Scaling laws are coming for real world robots as well:
…Which means robots are about to get really, really good, really, really quickly… 

Scaling laws: They also find scaling laws - "training on more trajectories reduces position tracking error, which is a positive signal", they write, and also note that "larger context windows produce better policies, which suggests that our generative policy performs a form of in-context adaptation that improves with scale." In general, "tracking error monotonically decreases with model size."
    Translation: Give us more data, bigger context windows, and more parameters in our model, and this will all get way better.