Simulation Workflow

In this section, we’ll complete the Unitree G1 static apple-to-plate workflow in Isaac Lab-Arena. The workflow covers environment setup and validation, OpenXR teleoperation data collection, GR00T 1.7 policy post-training, and closed-loop evaluation.

The task is a static, no-locomotion apple-to-plate task.

Teleoperation environment for this task.

The G1 stands in front of a shelf, uses its arms to pick up an apple, and places it onto a plate on the same shelf. WBC actively balances the standing pose, and the lower body does not walk, squat, or turn.

The training and evaluation steps use a standalone clone of NVIDIA’s Isaac-GR00T repository rather than the GR00T submodule pinned inside Arena.

Evaluation runs over Arena’s server-client remote-policy architecture: the GR00T server hosts the fine-tuned checkpoint in its own virtual environment, and Arena’s client runs the simulation in the standard Arena container and queries the server over ZeroMQ.

Task Overview

Task Specification (click to expand)

Property	Value
Task name	galileo_g1_static_pick_and_place
Tags	Tabletop manipulation, no locomotion
Skills	Pick, place; no walk, squat, or turn
Embodiment	Unitree G1, 29 DOF humanoid with WBC for balance only
Interop	LeRobot dataset format, converted from teleoperation HDF5
Scene	Galileo Lab Environment with a single shelf
Manipulated object	Apple rigid body
Destination	Clay plate on the same shelf
Policy	GR00T 1.7, fine-tuned through standalone Isaac-GR00T
Post-training	Imitation learning
Dataset	Self-recorded through teleoperation or nvidia/Arena-G1-Static-PickNPlace-Task
Checkpoint	Self-trained or nvidia/GN1x-Tuned-Arena-G1-Static-PickNPlace
Physics	PhysX, 200 Hz at 4 decimation
Closed-loop control	Yes, 50 Hz control
Metrics	Success rate

What We’ll Cover

Lesson	What You’ll Do
Sim Setup: Isaac Lab-Arena	Install Isaac Lab-Arena, prepare workflow directories, and run validation tests
Sim Environment Code Review	Review the environment registration, embodiment choices, object placement, and task success logic
Sim Teleop and WBC	Collect OpenXR teleoperation demonstrations with the G1 static task
Sim Data Export	Convert recorded HDF5 demonstrations to LeRobot format
GR00T Fine-Tuning on Sim Data	Fine-tune GR00T 1.7 from the standalone Isaac-GR00T checkout
Sim Evaluation	Run closed-loop policy inference through Arena’s server-client evaluation path

Learning Objectives

By the end of this Simulation Workflow section, you’ll be able to:

• Validate the galileo_g1_static_pick_and_place environment in Isaac Lab-Arena.

• Review the static apple-to-plate environment code path in Isaac Lab-Arena.

• Collect OpenXR teleoperation demonstrations for the static apple-to-plate task.

• Convert HDF5 teleoperation recordings to LeRobot format.

• Fine-tune GR00T 1.7 using a standalone Isaac-GR00T checkout.

• Evaluate the fine-tuned checkpoint in closed loop through Arena’s remote-policy architecture.

---