Train an ACT policy to make the highest pile possible

[dimos-bot]

Demos:

1. Ask it to make shapes like circles, lines etc - with agent input and arms execute

---

Synced from DIM-1477