Skip to main content

Resource Download


This is the official technical documentation for Global Humanoid Robot Challenge 2026 (GHRC 2026). It is built on the LeRobot framework and provides an end-to-end workflow for a humanoid robot simulation platform, covering physics simulation, data collection, model training, and deployment.

1. Project Overview

This documentation is intended for GHRC 2026 participants and R&D teams, providing a unified baseline implementation:

  • Build a high-fidelity humanoid robot simulation environment based on NVIDIA Isaac Sim
  • Collect data via keyboard teleoperation and export in the standardized LeRobotDataset V3.0 format
  • Train and fine-tune models using imitation learning algorithms (e.g., ACT, Pi0)
  • Reproduce and benchmark quickly with official pretrained weights

2. Key Capabilities

CapabilityDescription
Simulation EnvironmentHigh-fidelity Walker S2 robot simulation based on NVIDIA Isaac Sim, supporting a 20-dimensional state space (14 arm joints + 4 gripper joints + 2 gripper control commands).
Data CollectionSupports keyboard teleoperation; exports LeRobotDataset V3.0 format.
Model TrainingSupports imitation learning algorithms such as ACT and Pi0.
4-View Real-time DisplaySupports real-time preview from 4 RGB cameras (head_left, head_right, wrist_left, wrist_right).

3. Resources

Some large files in this project are hosted on Hugging Face. Please download them before first use:

Resource TypeLocal DirectoryRemote
🤖 Simulation environment & robot assetsassets/ (Git submodule)UBTECH-Robotics/challenge2026_assets
📊 Training datasetdatasets/UBTECH-Robotics/challenge2026_dataset
MinimumRecommendedIdeal
OSUbuntu 22.04 / 24.04; Windows 10 / 11Ubuntu 22.04 / 24.04; Windows 10 / 11Ubuntu 22.04 / 24.04; Windows 10 / 11
CPUIntel Core i7 (7th Gen); AMD Ryzen 5Intel Core i7 (9th Gen); AMD Ryzen 7Intel Core i9 (X-series or higher); AMD Ryzen 9 / Threadripper (or higher)
Cores4816
RAM32GB64GB64GB
Storage50GB SSD500GB SSD1TB NVMe SSD
GPUGeForce RTX 4080GeForce RTX 5080RTX PRO 6000 Blackwell
VRAM16GB16GB48GB
DriverLinux: 580.65.06; Windows: 580.88Linux: 580.65.06; Windows: 580.88Linux: 580.65.06; Windows: 580.88

We recommend using larger RAM and VRAM capacities, especially for model training. Also, if you installed the 595 driver, Isaac Sim may crash inside the Docker container later; therefore, we recommend using driver version 580.

3.2 Tool Requirements

ToolVersionNotes
CUDA12.8Official Guide
DockerlatestOfficial Guide
NVIDIA Container ToolkitlatestOfficial Guide
Hugging Facelatestpip install huggingface-hub; huggingface-cli --help (verify installation)
Gitlatestsudo apt update; sudo apt install git -y; git --version (verify version)
MinicondalatestOfficial Guide (optional)