Environment Variable Reference#

Environment variables select external runtimes, assets, and services for the simulated G1, simulated H2, and physical G1 workflow paths. The tables also list the shell variables that carry a run’s files between stages.

Starting a New Terminal#

Restore the settings for the stage you are resuming on the machine where it runs. Use the same asset, dataset, and checkpoint paths as the preceding stage. For an existing H2 run, retain H2_RUN_DIR; creating a new timestamped directory would disconnect the later commands from your recording.

Export variables read by a launched process with export NAME=value. Shell variables are expanded by the command examples and need an assignment in each terminal that uses them, even when the guide originally exported them. Shell state does not transfer to a new terminal, another machine, or an already running container.

Stage

Restore

G1 simulation and recording

Simulation runtime; CloudXR setup from Installation

H2 simulation, recording, replay, and generation

Simulation runtime and H2 assets; CloudXR in the separate server terminal

G1 or H2 optional data generation

AutoData overrides, if provisioning used non-default paths; the selected run’s input/output variables

G1 or H2 dataset loading, training, and evaluation

GR00T checkout, matching run variables; H2 assets for H2 simulation/evaluation

Physical G1 teleoperation, recording, and deployment

Isaac ROS workspace and activation; source the CloudXR environment in each teleoperation terminal

Simulation Runtime#

Variable

Requirement and Value

Setup

ACCEPT_EULA, OMNI_KIT_ACCEPT_EULA

For non-interactive Isaac Sim launch, Y and YES respectively, only after accepting the license terms.

Installation

E2E_ARENA_ROOT

Optional override for the pinned Isaac Lab-Arena checkout. Unset uses the managed .arena/ checkout under this repository.

How Arena Is Obtained

ISAAC_GR00T_ROOT

Absolute path to the clean, synced Isaac-GR00T checkout pinned in runtime.toml. Required by the documented training, loader-validation, and evaluation commands. Launchers also accept --gr00t-root.

Prepare the GR00T Checkout

The EULA settings apply to simulation processes, including replay and generation. They do not accept the separate CloudXR license.

H2 Assets#

Restore these in each terminal that starts H2 simulation. The source contracts are configs/robots/h2.yaml and configs/robots/h2_pink.yaml.

Variable

Requirement and Value

Setup

E2E_H2_ASSET_CACHE

Asset cache root used during preparation. The runtime default is $HOME/.cache/e2e-reference-app/h2plus. Export it when using another location, and always for sim-arena-h2, even at the default path.

Prepare the Simulation Assets

E2E_H2_PINK_URDF

Required for the Pink recording, replay, and generation embodiment. Use the pink_urdf path printed by the sidecar materializer.

Prepare the Simulation Assets

E2E_H2_AGILE_POLICY

Required H2 AGILE ONNX file. Its digest must match the robot configuration; a filename alone is insufficient.

Download the AGILE Policy

The H2 validation command exercises both Pink and joint embodiments and needs all three settings plus ISAAC_GR00T_ROOT.

H2 CloudXR Server#

Set these in the CloudXR terminal, separate from the recording terminal. Start CloudXR provides the client download and launch commands.

Variable

Requirement and Value

TELEOP_WEB_CLIENT_STATIC_DIR

Directory containing the downloaded pinned index.html and bundle.js for --host-client. Restore the same directory when restarting the server.

Isaac AutoData#

These overrides are optional. Restore any that you used during provisioning. tools/launch_autodata.sh owns their defaults; explicit command-line options take precedence where provided.

Variable

Default and Use

ISAAC_AUTODATA_ROOT

The sibling Isaac-AutoData-e2e directory. Selects the pinned clean checkout; also settable with --autodata-root.

AUTODATA_PYTHON

Interpreter in the isaac_autodata Conda environment. Override with an absolute Python path or --autodata-python. Native operations and validate-dataset use this interpreter; Arena generation uses the repository’s Arena runtime.

AUTODATA_CONDA_ENV

isaac_autodata. Selects the Conda environment when AUTODATA_PYTHON is unset.

E2E_AUTODATA_OUTPUT_DIR

outputs/autodata/ under this repository. Used when output paths are not supplied explicitly.

E2E_AUTODATA_GENERATION_TIMEOUT_SECONDS

420 seconds. Whole-process generation deadline, overridden by --generation-run-timeout-seconds; use the H2 timeout guidance for larger runs.

Run and Artifact Shell Variables#

These names belong to the guide’s shell commands. The applications receive their values through file paths and command-line arguments. Recreate them for the existing run before copying commands from a later stage.

Variable

Value to Restore

Guide

E2E_ROOT

Absolute path to this repository, captured before changing into Isaac-GR00T.

G1 training, H2 training

ARENA_ROOT

Output of ./tools/fetch_arena_checkout.sh, passed as --arena-root. This shell name does not replace the E2E_ARENA_ROOT launcher override.

G1 evaluation, H2 evaluation

TRAIN_DATASET, MODALITY_CONFIG

G1 LeRobot dataset and G1 modality configuration from the training guide.

G1 training

TRAIN_OUTPUT

Training output directory for the selected G1 or H2 run.

G1 training, H2 training

G1_CHECKPOINT, H2_CHECKPOINT

Selected published model directory or completed checkpoint directory from the matching robot’s training run. G1 also supports the released checkpoint described in its evaluation guide.

G1 evaluation, H2 evaluation

G1_CHECKPOINT_SOURCE, H2_CHECKPOINT_SOURCE

published for the downloaded workflow model or trained for your own checkpoint. For the separate G1 released checkpoint, use released. Keep the source paired with the selected checkpoint path.

G1 evaluation, H2 evaluation

H2_RUN_DIR

Absolute path to the existing outputs/h2/<timestamp>/ directory. Choose a new directory when starting training with the published dataset.

H2 recording, H2 training

H2_RECORDING, H2_SOURCE_DEMOS

Recorded source HDF5 and its expected accepted episode count.

H2 recording

H2_GENERATED, H2_TARGET_DEMOS

Generated HDF5 and requested successful generation count.

H2 generation

H2_GENERATION_TIMEOUT_SECONDS

Measured wall-clock budget passed as --generation-run-timeout-seconds. This shell value overrides the launcher’s environment default through that argument.

H2 generation

H2_TRAINING_HDF5, H2_TRAINING_DEMOS

Selected source or generated HDF5 and its expected episode count. Keep the path and count paired. For published-dataset loader validation, set H2_TRAINING_DEMOS=1000; no HDF5 input is needed.

H2 conversion

H2_LEROBOT, H2_MODALITY

Converted or downloaded dataset directory and H2 modality configuration.

H2 conversion, H2 training

H2_SMOKE_OUTPUT

Separate output directory for the short training check.

H2 training

H2_AGILE_REPOSITORY, H2_AGILE_REVISION, H2_AGILE_FILENAME

Temporary download values read from configs/robots/h2.yaml. Re-read them for a download; later stages use E2E_H2_AGILE_POLICY.

H2 policy download

DATASET_DIR

Physical G1 LeRobot output directory on the machine doing conversion. Transfer the complete directory when training elsewhere.

Physical G1 conversion

Physical G1#

The physical path uses the Isaac ROS environment on Thor. Its x86_64 conversion, training, and export stages use the environments stated in their own guides.

Setting

Requirement and Value

Setup

ISAAC_ROS_WS

Absolute workspace path on the current machine. Setup persists $HOME/workspaces/isaac_ros-dev/ in the host’s shell configuration. Verify that the same workspace is visible inside the container.

Create the Isaac ROS Workspace

LANG, LC_ALL

UTF-8 locale; setup persists en_US.UTF-8 for both on the Thor host and exports LANG in the setup terminal.

Initialize the Isaac ROS CLI

Isaac ROS and ROS_DISTRO

Run isaac-ros activate in each new terminal used for ROS commands. Use its configured ROS environment.

Physical G1 teleoperation

CloudXR environment

With CloudXR running, source ~/.cloudxr/run/cloudxr.env inside each Isaac ROS terminal used to launch teleoperation. Use the generated file rather than copying individual service variables.

Initial Test in MuJoCo

CUDA_HOME, CUDA_VISIBLE_DEVICES, PYTORCH_CUDA_ALLOC_CONF

The training example supplies /usr/local/cuda, GPU 0, and expandable_segments:True for that command only. Match the toolkit path and GPU selection to the training workstation.

Launch Training

The physical training and LEAPP export guides pass dataset, checkpoint, and bundle paths explicitly. They do not use the simulation launchers’ GR00T or H2 environment variables.