Unitree G1 Real-Robot Static Apple Pick-and-Place#
This workflow path covers a Unitree G1 humanoid with Dex3-1 hands, a head-mounted Intel RealSense camera, and a Jetson AGX Thor connected externally or through the Unitree Thor backpack. Thor runs Isaac ROS; AGILE whole-body control stabilizes the lower body while an operator or GR00T policy commands the upper body for an apple-to-plate manipulation task.
AVAILABLE
Unitree G1 · Isaac ROS · Jetson AGX Thor · GR00T N1.7
Warning
Read Unitree G1 Introduction and Safety in full before powering on the robot. Follow Unitree’s operating documentation and your organization’s robot-safety procedures. Use at least two people: one to operate and one to monitor safety.
Understand the Workflow#
Understand each stage of the reference workflow and how the physical data path differs from the simulation path.
Review where physical data collection fits in the pipeline and which data formats and task settings each dataset and checkpoint requires before running the stages.
Follow the Workflow Path#
Complete the stages in order. Each page states where it runs and what it produces. Recording uses the Isaac ROS MCAP recorder on Thor. The native and Arena HDF5 recording modes apply to simulation paths.
Use the Environment Variable Reference when restoring the workspace and service environment in a new terminal.
1
Review the safety checkpoints for teleoperation, recording, and deployment. Prepare the workspace and operating roles before connecting the robot.
2
Install the isaac-ros CLI, ROS 2 workspace, source repositories, container
environment, and RealSense support on the computer connected to the Unitree G1.
3
Verify the XR headset and CloudXR session in simulation, then run whole-body XR teleoperation on real hardware through the ROS safety controller.
4
Capture teleoperation demonstrations in MCAP format, keeping only clean runs that follow the recording guidance for GR00T training.
5
Install the converter, merge one or more recording sessions into a single LeRobot dataset, and verify its output paths and metadata.
6
Validate the dataset and modality configuration, fine-tune GR00T N1.7 on the real-world dataset, and locate the checkpoint for export.
7
Package the fine-tuned policy into a deployable LEAPP bundle that runs from C++ with its sensor inputs and command outputs already connected.
8
Deploy the bundled apple-to-plate policy to Thor, replace it with your own export, and measure real-world success rate.
Hugging Face Dataset and Checkpoint#
Skip collecting or fine-tuning your own demonstrations with the published physical-robot artifacts already linked from the conversion and fine-tuning pages. These Hub pages may require accepting NVIDIA access terms.
DATASET ON HUGGING FACE
nvidia/GR00T-N1.7-AppleToPlate
LeRobot demonstrations for the Unitree G1 real-robot static-apple workflow.
MODEL ON HUGGING FACE
nvidia/GR00T-N1.7-ApplePnP-V1
Fine-tuned GR00T N1.7 checkpoint for the same physical Unitree G1 task.
Supporting Guidance#
Diagnose RealSense, CloudXR, ROS domain, and Thor throttling issues that block teleoperation, recording, or deployment.
Testing and Limitations#
The repository’s automated tests do not run on the physical robot. The setup requirements specify JetPack 7.1 / Isaac ROS 4.5 and explain why the exact v0.1 installation cannot be fully reproduced. The historical 68% result also lacks the original scoring criteria needed for direct comparison.