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

[Gumin-Research]

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.

[robomantri]

Hey — just getting into G1 + Inspire RH56DFTP + GR00T N1.7. Saw Whu-BrainLab did something similar with N1.5: https://github.com/Whu-BrainLab/Embodied-AI-Tutorial
https://mp.weixin.qq.com/s/U-b_yLjKwsUXidBHCxYeew
If anyone's actively working this stack and wants to compare notes on hand schema mapping or data collection — I'm down.

[rylinjames]

on your specific questions:

the REAL_G1 embodiment in N1.7 uses left_hand/right_hand action keys but the pretraining data almost certainly used Dex3 or fakehand configs, not Inspire RH56DFTP. NVIDIA hasn't published the hand-specific breakdown of pretraining data. zero-shot with correct obs/action schema mapping is worth trying for arm/torso tasks but expect hand manipulation to fail or produce garbage until you finetune.

the NEW_EMBODIMENT path is the right call for Inspire. the 20-40 demo number from the N1.6 blog still seems to hold for N1.7 for simple pick/place but you'll want more for dexterous contact-rich tasks with the 24-joint hand.

for the Isaac Lab to LeRobot v2 dataset pipeline: UnitreeG1RetargeterCfg outputs in Isaac Lab's native format. you'll need to convert to LeRobot v2 episode format (HDF5 with specific key naming) before passing to GR00T finetuning. there's no official converter yet -- a few people have written custom scripts, worth checking the GR00T discord for the latest.

[toyoxx]

@robomantri @Gumin-Research I am working on the exact same stack, should we just create a Discord and share ideas/compare notes? I have some demos working with an Inspire Python wrapper and the decoupled gr00t policy, but my next step is definitely going towards actually using the Whole Body pipeline with the inspire hands. Unitree has a C++ wrapper for the Inspire hands, and I think looking at their teleop stack (https://github.com/unitreerobotics/xr_teleoperate) might give us some insight on how to integrate them in the NVIDIA teleop pipeline.

[robomantri]

sure hit me up

[mattiafornari]

Hi everyone,

I am planning a precision manipulation task (tightening a screw) using the Unitree G1 EDU U6 equipped with Inspire Hand RH56DFTP. Since I don't have real-world datasets, I intend to use a Meta Quest 2 for data collection, but first I need to figure out the training infrastructure and asset readiness.

I'm looking for some guidance from the community and the NVIDIA team regarding my planned workflow:

1. G1 + Inspire Hand USD Assets & Sensors

• Is there a verified or community-shared G1 + Inspire Hand USD asset available that already integrates depth cameras, LiDAR, and tactile sensors of the hands?
• If I need to build/modify it myself, what is the recommended workflow within Isaac Lab to properly map and simulate the 24-DoF kinematics and tactile sensor arrays of the Inspire Hand and the depth camera and lidar?

2. Cloud Fine-Tuning Feasibility (Brev.dev)

• Is GR00T N1.7 fully compatible with cloud-native orchestration on Brev.dev? Do you provide official Docker containers or templates optimized for this platform?
• Since the Inspire Hand might not be natively covered in the N1.7 pre-training weights, is it realistic to expect successful fine-tuning for contact-rich manipulation tasks using purely synthetic data generated via Isaac Lab in the cloud?

3. Sim-to-Sim Validation (MuJoCo) & Local Hardware

My local setup is limited to an RTX A4000 (16GB VRAM). My plan is to offload the heavy Isaac Lab training to Brev.dev, but run local Sim-to-Sim validation in MuJoCo before moving to the physical robot.

• Is there a supported path to export policies trained in Isaac Lab/GR00T into a MuJoCo-compatible environment for inference testing?
• Will 16GB of local VRAM be enough to run these Sim-to-Sim evaluation loops for the G1 + Inspire Hand model?

Any tips, repository links, or documentation regarding this hybrid cloud/local pipeline would be super helpful!

Thanks!

[Kaweees]

@robomantri @Gumin-Research I am working on the exact same stack, should we just create a Discord and share ideas/compare notes? I have some demos working with an Inspire Python wrapper and the decoupled gr00t policy, but my next step is definitely going towards actually using the Whole Body pipeline with the inspire hands. Unitree has a C++ wrapper for the Inspire hands, and I think looking at their teleop stack (https://github.com/unitreerobotics/xr_teleoperate) might give us some insight on how to integrate them in the NVIDIA teleop pipeline.

Could you add me as well? Would be interested in sharing notes

[robomantri]

@Kaweees @toyoxx https://discord.gg/pweePvF2wB I have created a discord

[mattiafornari]

@robomantri Can I join the Discord too?

[robomantri]

@mattiafornari Sure