Prerequisites

Before starting either workflow, verify that you have the required background knowledge and the correct hardware and software for your chosen path.

Recommended Skills

All tools run on Linux and use Docker, so you should be comfortable navigating directories, running scripts, reading terminal logs, and working with Docker containers.

This course assumes intermediate Python proficiency. You run training scripts, modify configuration files, and inspect data throughout both workflows.

Humanoids are highly sophisticated robots. This course assumes you have a basic understanding of their operation and control, along with some prior experience operating real robots.

Tip

To build foundational knowledge before starting, refer to the Physical AI learning offerings and resources.

Shared Requirements (Both Workflows)

Both paths collect demonstrations through teleoperation, convert them to LeRobot format, and post-train GR00T 1.7. The requirements below apply whether you work in simulation or on the physical G1.

Activity	Requirement
Teleoperation	Any Isaac Teleop supported input device such as Meta Quest 3 or PICO 4 Ultra, except Apple Vision Pro, plus a workstation and network that meet the Isaac Teleop system requirements
Training	NVIDIA GPU with at least 48 GB VRAM for GR00T 1.7 post-training (1 GPU minimum; 8× GPUs recommended for the real-robot workflow at scale)

Simulation Workflow Requirements

Workstation

You need a workstation capable of running Isaac Sim with real-time physics and rendering.

These requirements follow the Isaac Sim 6.0 hardware requirements for x86_64 workstations.

Component	Minimum	Recommended	Ideal
OS	Ubuntu 22.04 or 24.04	Ubuntu 22.04 or 24.04	Ubuntu 22.04 or 24.04
CPU	Intel Core i7 7th Gen or AMD Ryzen 5	Intel Core i7 9th Gen or AMD Ryzen 7	Intel Core i9, Intel X-series or higher, AMD Ryzen 9, Threadripper or higher
CPU cores	4 cores	8 cores	16 cores
RAM	32 GB	64 GB	64 GB
Storage	50 GB SSD	500 GB SSD	1 TB NVMe SSD
GPU	GeForce RTX 4080	GeForce RTX 5080	RTX PRO 6000 Blackwell
VRAM	16 GB	16 GB	48 GB
Driver	Linux 580.65.06	Linux 580.65.06	Linux 580.65.06

Note

The Isaac Sim container runs on Linux. Loading Isaac Sim assets and running some extensions require internet access. Isaac Sim does not support GPUs without RT Cores, such as A100 or H100.

Isaac Lab and GR00T workflows can require additional RAM, VRAM, and storage beyond the base Isaac Sim requirements, especially for training or large scenes.

Software

Install ffmpeg on the simulation workstation before collecting or converting demonstration data. The simulation workflow uses it when writing and processing video observations.

sudo apt-get update
sudo apt-get install -y ffmpeg

Real Robot Workflow Requirements

Component	Details
Robot	Unitree G1 connected to Thor using Ethernet, with Dex3-1 hands
Camera Extension	USB3 extension cable long enough to route the head-mounted RealSense camera USB connection from the G1 head to Jetson AGX Thor
Edge Compute	Jetson AGX Thor Developer Kit for real robot setup, teleoperation, recording, and deployment. The developer kit includes a Jetson T5000 module with a 2560-core NVIDIA Blackwell GPU, 14-core Arm Neoverse-V3AE CPU, 128 GB LPDDR5X memory, 1 TB NVMe storage, USB3 ports, and 5 GbE RJ45 networking.

Refer to Shared Requirements (Both Workflows) for teleoperation device and training GPU requirements.

Workspace Requirements

The physical workspace for the real robot workflow requires controlled conditions for reliable policy performance.

Element	Specification
Robot	Unitree G1, standing in front of the table, facing the table, using built-in RealSense camera
Table	Standard height table with a black tablecloth
Apple	Red, placed on the table to the left of the plate, within left arm reach
Plate	White, placed on the table within left arm reach
Background	White wall behind the table
Lighting	Consistent indoor lighting, avoid strong shadows or glare