Getting enough real-world demonstrations to train a humanoid is expensive and slow, and this build tackles that from a different angle: the maker built a self-service tea station as a fully simulated scene and sent a Unitree G1 humanoid in to brew tea inside it. The idea being explored is whether environments like this can act as a source of synthetic demonstrations, with the trajectories collected for imitation learning.
The setup itself is nicely concrete. There is a wooden countertop and a tiered rack holding black ceramic mugs, plus an electric kettle, a glass jar, and a condiment organizer tray stocked with tea bags and spoons. The G1 walks to the rack, picks up a mug, and drops a tea bag in, all across five scripted task phases. The movements were scripted separately and then run in Isaac Sim, with PhysX handling the physics, including fluid dynamics for the pouring. The pour came out remarkably clean: 147 ml went in and only 1.9 ml spilled.
Two details stand out to me. First, the per-task playback speeds vary a lot, from 1.5x up to 4.4x, which seems a sensible way to keep the footage watchable. Second, the maker is upfront about the rough edges: keeping the G1 stable during interactions took significant work, and the run shows imperfect contacts and slips, with a few mm of mug slip visible in the hand. The recorded heel strike of 2.0 mm is also noted. The big open question, which the maker flags directly, is what imitation learning experts make of this whole approach.