[REAL_G1 / N1.7] Is there an official Isaac Sim setup and is zero-shot behavior expected?

Summary

We are trying to run the nvidia/GR00T-N1.7-3B base checkpoint with the provided REAL_G1 embodiment in an Isaac Sim Unitree G1 environment.

The model returns valid 40-step action trajectories, but different prompts do not result in recognizable task-directed behavior. For example, neither:

Pick up the apple and place it on the plate.

nor:

Raise your right arm straight up above your head.

produces the corresponding visible motion. The robot only makes small balance-like movements.

Questions

Is there an official executable Isaac Sim or Isaac Lab reference setup for nvidia/GR00T-N1.7-3B using REAL_G1?

Is the N1.7 base checkpoint expected to produce recognizable zero-shot behavior with REAL_G1, or does REAL_G1 only provide the correct modality structure for G1 post-training and therefore still require task-specific demonstrations?

Short setup description

We use IsaacLab-Arena with its Unitree G1 whole-body controller and the official GR00T policy server and PolicyClient. We added the required observations, including the [-20, 0] video history and pelvis-relative wrist poses in XYZ_ROT6D format.

We also validated the downstream action path by replacing the model output with a deterministic right_arm trajectory using the same shape and datatype as the GR00T output. This caused the simulated G1 to raise its arm visibly above its head. Therefore, the joint-action mapping and controller path are functional, but the GR00T prompts tested so far do not produce recognizable task-directed motion.

Any official simulation example or clarification of the expected zero-shot capabilities would be greatly appreciated.