Synthetic Data¶

This chapter is the Isaac Sim datagen pipeline: USD scenes → ROS 2 interface → task-queue orchestration → LeRobot recording / export.

This How-To category is the Isaac datagen pipeline, not Gazebo or teleop. Reference pages stay the source of APIs; this chapter only explains how the pieces connect.

Documented branches

robot_action_composer @ feature/dex-grasp-generator
lerobot_ros2 @ feature/sim-grasp-datagen

Those branches are newer than main. Follow them, not main.

Terminology

Humanoid here means a wheeled-arm humanoid (mobile base + arms, e.g. FiveAges W2). It does not mean a bipedal or footed humanoid.

Pipeline¶

FaSim-Isaac + robot_usds (+ env / private USD)
        ↓  Isaac scene, ROS 2 bridge
ros2_robot_interface
        ↓  motion / state / gripper / Nav2 client
robot_action_composer  (task_queue YAML + skills)
        ↓  StageTarget sequences via ROS2RobotInterface
lerobot_ros2  (record_datasets.py → LeRobot dataset on disk)

Stage	Role in datagen	Start here
FaSim-Isaac	Isaac Sim line for synthetic scenes. Scripts pull USD and start Sim.	Isaac scenes and USD
robot_usds	Robot USD superproject used by FaSim (`robots/`).	same page
ros2_robot_interface	Python interface used by motion and recording.	Interface and orchestration
robot_action_composer	`task_queue` YAML + skill registry; `motion-generation` / `ros2-stack`.	same page
lerobot_ros2	Monorepo that hosts the Isaac example workspace and recording plugins. Stop at record / export.	LeRobot record and export

In this section¶

Baseline¶

ROS 2 Jazzy
Python 3.12 (same baseline as the lerobot_ros2 / composer READMEs on those branches)

Out of scope¶

This chapter does not cover:

Gazebo
ros2-viser or fa-py-libraries
Policy training, online inference, or dataset consumption (examples/IsaacSim/policy_training/train.py, inference.py)

Package-level APIs: ros2_robot_interface, robot_action_composer, lerobot_ros2, FaSim-Isaac, robot_usds.