Unitree G1 Static Apple Pick & Place#
The simulated Static Apple Pick & Place workflow path uses a Unitree G1 to move an apple to a plate in NVIDIA Isaac Lab. It carries those demonstrations through GR00T N1.7 training and closed-loop evaluation.
AVAILABLE
Understand the Workflow#
Understand the role of each stage and the difference between simulation and physical data collection.
Review the pipeline, robot-control software, and data requirements before running the stages.
Before You Start#
Complete Installation to prepare the workstation and simulation environment. Recording demonstrations requires an XR headset; check the XR headset, handheld controller, and network requirements before starting the recording stage.
Use the Environment Variable Reference when restoring runtime and run settings in a new terminal.
Follow the Workflow Path#
Follow the five stages in order, or use the
published stage inputs
to start at conversion, training, or evaluation. After recording, optional
Isaac AutoData generation can expand
your demonstrations before conversion. The recording guide uses Isaac Lab-Arena
with Pink IK for upper-body control and AGILE for lower-body control
(g1_wbc_agile_pink). Native Isaac Lab with AGILE is a supported alternative;
both provide the command fields required by the supplied training configuration.
1
Launch the scene and practice driving the Unitree G1 with AGILE.
2
Record HDF5 demonstrations in Isaac Lab-Arena with Pink IK and AGILE using a supported XR device. The task scores each episode automatically.
3
Convert the HDF5 data to LeRobot and validate the fields required by the policy training modality for Unitree G1.
4
Train GR00T N1.7 on the validated dataset of Unitree G1 demonstrations.
5
Run the trained policy in the workflow’s scene and collect episode metrics.
Workflow Artifacts#
Commands use outputs/g1/ under the repository root for one run. Choose new
recording and output paths when starting another run.
Stage |
Output |
Next Input |
|---|---|---|
Recording in Isaac Lab-Arena with Pink IK and AGILE |
|
Direct conversion or Arena AutoData generation |
Optional Arena generation |
|
Arena conversion |
Conversion |
|
GR00T training dataset |
Training |
|
Checkpoint preflight and policy server |
Evaluation |
|
Episode metrics and HTML report |
Keep the original manifested recordings. Merged HDF5 files cannot be used for replay or AutoData with the current merger.
For the native AGILE alternative, the guides use recording-native.hdf5 and
generated-native.hdf5. Native fixed-base data requires a custom training and
evaluation configuration because it lacks navigation and base-height commands.
Tip
Start conversion with a generated HDF5 download. LeRobot dataset and model downloads are included in the G1 policy training and G1 policy evaluation steps.
Hugging Face Dataset and Checkpoint#
The checkpoint below is separate from the released checkpoint pinned in
runtime.toml.
DATASET ON HUGGING FACE
nvidia/GR00T-E2E-Reference-Workflow-SIM-G1-N1.7-AppletoPlate-Dataset
Recorded HDF5, generated HDF5, and LeRobot demonstrations for the Unitree G1 static-apple simulation workflow.
MODEL ON HUGGING FACE
nvidia/GR00T-E2E-Reference-Workflow-SIM-G1-N1.7-AppletoPlate-Model
Fine-tuned GR00T N1.7 checkpoint for the same Unitree G1 simulation task.
Supporting Guidance#
Optional: turn a few AGILE recordings into many with Isaac AutoData. The launcher selects the recorded action layout and manipulating hand automatically.
Testing and Limitations#
This workflow covers one Unitree G1 task on Linux x86_64. By default, recording
and checkpoint evaluation use the same success criteria.
Manually ending a recording does not change its success label. Pass
--no-auto_success to let the operator decide which episodes to accept. The
released checkpoint can test whether the components work together; its results
do not establish expected performance in this workflow’s scene.
See Reference for supported configurations and limitations.