Zero-shot deployment of GR00T N1.7 on Unitree G1 + Inspire RH56DFTP hand

Question about Inspire hand support and zero-shot deployment

First of all, thank you for the amazing work on GR00T N1.7 — the release and documentation have been extremely helpful for getting started quickly. I'm currently evaluating the model on my setup and have a few questions I'd love some guidance on.

My Setup

• Robot: Unitree G1 EDU Ultimate (29 DoF body)
• Hand: Inspire RH56DFTP 5-finger dexterous hand (24 joints total, 17 tactile sensors per hand)
• Simulation: Isaac Lab Isaac-PickPlace-G1-InspireFTP-Abs-v0 task (Pink IK controlled, fixed base)
• Hardware: H100 80GB

Questions

1. Inspire RH56DFTP in pretraining data Has the N1.7 base model been exposed to data from the Inspire RH56DFTP 5-finger hand during pretraining? Looking at the REAL_G1 embodiment (real_g1_relative_eef_relative_joints), the state/action includes left_hand / right_hand keys, but it's unclear which specific hand configuration was used. Was it Dex3-1, Inspire, or a mix?

2. Hand-specific embodiment tags Are there any plans to release embodiment tags or finetuned checkpoints specifically targeting G1 + Inspire RH56DFTP? Or is the recommended path to use NEW_EMBODIMENT and finetune from scratch?

3. Zero-shot feasibility Realistically, how feasible is zero-shot deployment of the base N1.7 model on a G1 + Inspire setup, assuming the observation/action schema is mapped correctly? Given the hand joint count and kinematics differ between hand variants, I'd expect zero-shot to produce poor results on contact-rich tasks, but I'm curious about NVIDIA's view.

4. Data collection recommendation If finetuning is the right path, what's the recommended minimum data scale for adapting to a new hand configuration on an existing robot body (G1)? The N1.6 blog mentioned 20-40 demos being sufficient for some tasks — is that still the case for hand-specific adaptation with N1.7?

5. Teleop data format For collecting teleop demonstrations on G1 + Inspire in Isaac Lab, is there a recommended pipeline that produces data directly in the LeRobot v2 format expected by the finetuning scripts? I see UnitreeG1RetargeterCfg in Isaac Lab's teleop setup, but it's not clear how to export to GR00T-compatible datasets.

Thanks again for the great release! Any pointers would be much appreciated.