Installation#

Choose a workflow path before installing its environment. This page prepares the workstation for simulated Unitree G1 and Unitree H2.

For physical Unitree G1, follow the physical workflow path. It provides separate Isaac ROS setup on Jetson AGX Thor and workstation setup for training and export. The simulation environment below is not required for that path.

System Requirements#

The simulation baseline follows the Isaac Sim 6.0 requirements. The repository’s checked runtime supports Linux x86_64.

Component

Minimum or required baseline

Recommended for this workflow

OS and architecture

Ubuntu 22.04 or 24.04, x86_64

Ubuntu 24.04, x86_64

CPU

4 cores

8 or more cores

System RAM

32 GB for simulation

64 GB for simulation; 128 GB or more for fine-tuning

Simulation and recording GPU

NVIDIA RTX GPU with RT cores and at least 16 GB VRAM; upstream lists NVIDIA RTX 4080

NVIDIA RTX 3090 is validated for the Unitree G1 stages except training; NVIDIA RTX 6000 Ada provides more memory for combined simulation and policy workloads

Fine-tuning GPU

48 GB VRAM for this repository’s tested training recipes

NVIDIA RTX 6000 Ada (48 GB); the Unitree H2 recipe is documented on NVIDIA RTX 5880 Ada (48 GB)

Linux NVIDIA driver

Use 580.95.05 as the Isaac Sim 6.0 tested baseline

A production driver qualified for your GPU and Isaac Sim version

CUDA runtime

CUDA 12.8 for the repository’s pinned PyTorch build

Keep the versions installed by uv.lock

Storage

50 GB SSD for the upstream simulator baseline

At least 500 GB SSD for environments, assets, datasets, and checkpoints

Python

3.12, installed by uv

Use the repository-local environment

The driver number is an upstream-tested baseline, not a measured lower support limit for this application. A100 and H100 GPUs lack RT cores and cannot run the simulator, even when suitable for a separate training server. The 16 GB simulator baseline does not guarantee enough memory to run a policy server concurrently. See the policy training configuration for Unitree G1 and policy training configuration for Unitree H2 for recipe-specific settings.

The locked environment supplies CUDA user-space packages; a separate system CUDA toolkit installation is not required for this repository’s wheel-based runtime. The host NVIDIA driver is still required. nvidia-smi reports the driver’s CUDA capability, which can be newer than the runtime used by PyTorch.

Inspect the machine before downloading. Check the OS and architecture:

cat /etc/os-release
uname -m

Check the CPU, memory, and available disk space:

lscpu
free -h
df -h .

Check the GPU and driver:

nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv

Network access is required to the source project, Python package indexes, GitHub, simulator assets, and Hugging Face. Training requires access to the gated model repositories described in Model Access. For XR recording, also prepare the headset and network before starting the recording stage.

XR Recording Requirements#

XR recording with NVIDIA Isaac Capture requires a Linux x86_64 workstation and a supported headset on the workstation’s LAN. Meta Quest 3 and PICO 4 Ultra can use the browser client through NVIDIA CloudXR. If a firewall is enabled, allow TCP ports 49100 and 48322 and UDP port 47998.

For Unitree G1, use XR hand tracking, handheld XR controllers, or Manus/Vive. Keyboard and SpaceMouse control are not supported for recording the static-apple task. Unitree H2 uses PICO 4 Ultra handheld XR controllers and requires Isaac Lab-Arena; follow the Unitree H2 Plus workflow.

After installing the simulation environment below, follow Unitree G1 XR setup to start CloudXR and connect the headset for a recording session.

Host Tools#

Install Git and uv using their installation instructions before cloning the repository.

Clone the Repository#

These instructions select main.

Choose a workspace with enough free space; the example uses ~/robot-workflows. Run each command below in order in the same terminal.

mkdir -p ~/robot-workflows
cd ~/robot-workflows
git clone --branch main \
  https://github.com/nvidia-isaac/GR00T-E2E-Reference-Workflow.git
cd GR00T-E2E-Reference-Workflow
git rev-parse HEAD

Keep the printed commit ID with your run results. To reproduce a specific revision, replace <tag-or-commit> below with a tag or commit available in the cloned GitHub repository:

git checkout --detach <tag-or-commit>

This repository has no Git submodules to initialize. Model weights and datasets are downloaded separately by their workflow stages.

For simulation-policy training, place the external Isaac-GR00T checkout alongside this repository, for example ~/robot-workflows/Isaac-GR00T. It has its own environment. Follow Prepare NVIDIA Isaac GR00T in Your Workspace to clone and select the exact revision recorded in runtime.toml; do not install GR00T into the simulator environment.

Create and Check the Simulation Environment#

From the GR00T-E2E-Reference-Workflow directory:

uv sync --frozen

This installs the simulator, host tools, and documentation dependencies into .venv. uv.lock specifies package versions; runtime.toml records the tested software versions and model revisions. No external Isaac Lab checkout or container is needed. Isaac Lab-Arena launchers fetch their pinned source checkout when needed. Optional AutoData provisioning is covered by its workflow stage.

After accepting the NVIDIA Isaac Sim license terms, run the canonical runtime preflight on the GPU workstation:

OMNI_KIT_ACCEPT_EULA=YES ACCEPT_EULA=Y \
  uv run --no-sync ./tools/validate_repository.sh runtime-preflight

Require a JSON result with "status": "ok". The check verifies runtime versions, CUDA availability, and the Pink IK solver dependencies. An "arena": "skipped" result means no external Arena checkout was supplied; the stage launcher checks that checkout when it is fetched. This preflight does not run a simulation or qualify a physical robot.

Choose a Workflow Path#