README
Global Humanoid Robot Challenge 2026 Baseline
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.
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, SmolVLA, Pi0)
Key Capabilities
| Capability | Description |
|---|---|
| Simulation Environment | High-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 Collection | Supports keyboard teleoperation; exports LeRobotDataset V3.0 format. |
| Model Training | Supports imitation learning algorithms such as ACT and Pi0. |
| 4-View Real-time Display | Supports real-time preview from 4 RGB cameras (head_left, head_right, wrist_left, wrist_right). |
Resources
Some large files in this project are hosted on Hugging Face. Please download them before first use:
| Resource Type | Local Directory | Remote |
|---|---|---|
| 🤖 Simulation environment & robot assets | assets/ (Git submodule) | UBTECH-Robotics/challenge2026_assets |
| 📊 Training dataset | datasets/ | UBTECH-Robotics/challenge2026_dataset |
Recommended Configuration
| Minimum | Recommended | Ideal | |
|---|---|---|---|
| OS | Ubuntu 22.04 / 24.04; Windows 10 / 11 | Ubuntu 22.04 / 24.04; Windows 10 / 11 | Ubuntu 22.04 / 24.04; Windows 10 / 11 |
| CPU | Intel Core i7 (7th Gen); AMD Ryzen 5 | Intel Core i7 (9th Gen); AMD Ryzen 7 | Intel Core i9 (X-series or higher); AMD Ryzen 9 / Threadripper (or higher) |
| Cores | 4 | 8 | 16 |
| RAM | 32GB | 64GB | 64GB |
| Storage | 50GB SSD | 500GB SSD | 1TB NVMe SSD |
| GPU | GeForce RTX 4080 | GeForce RTX 5080 | RTX PRO 6000 Blackwell |
| VRAM | 16GB | 16GB | 48GB |
| Driver | Linux: 580.65.06; Windows: 580.88 | Linux: 580.65.06; Windows: 580.88 | Linux: 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.
Tool Requirements
| Tool | Version | Notes |
|---|---|---|
CUDA | 12.8 | Official Guide |
Docker | latest | Official Guide |
NVIDIA Container Toolkit | latest | Official Guide |
Hugging Face | latest | pip install huggingface-hub; huggingface-cli --help (verify installation) |
Git | latest | sudo apt update; sudo apt install git -y; git --version (verify version) |
Miniconda | latest | Official Guide (optional) |
Technical Documentation Index
The complete baseline workflow consists of seven stages. We recommend following the documents in order; you may also jump directly to the stage you need based on your current progress.
| # | Document | Description |
|---|---|---|
| 1 | Resource Download | Project overview, key capabilities, hardware requirements, tool requirements, and HuggingFace resource links. |
| 2 | Environment Setup | Clone the repository, download simulation assets and datasets, build the Docker image, and configure keyboard evdev. |
| 3 | Quick Start | Start the runtime environment, keyboard teleoperation, data recording, dataset replay, and key mappings. |
| 4 | Model Training | Training guides for ACT, Diffusion Policy, π₀ (PI0), π₀.₅ (PI05), and SmolVLA with full hyperparameters. |
| 5 | Policy Inference | Run inference with a trained policy model and automatically record the results. |
| 6 | 4-Camera Real-time Display | Real-time preview from 4 RGB cameras and visualization configuration in teleoperation, recording, replay, and inference modes. |
| 7 | Automated Data Collection Tool | Collect task data through automated data collection scripts. |
Contributors
Similar repos
Languages
Python
99.7%
README
Global Humanoid Robot Challenge 2026 Baseline
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.
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, SmolVLA, Pi0)
Key Capabilities
| Capability | Description |
|---|---|
| Simulation Environment | High-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 Collection | Supports keyboard teleoperation; exports LeRobotDataset V3.0 format. |
| Model Training | Supports imitation learning algorithms such as ACT and Pi0. |
| 4-View Real-time Display | Supports real-time preview from 4 RGB cameras (head_left, head_right, wrist_left, wrist_right). |
Resources
Some large files in this project are hosted on Hugging Face. Please download them before first use:
| Resource Type | Local Directory | Remote |
|---|---|---|
| 🤖 Simulation environment & robot assets | assets/ (Git submodule) | UBTECH-Robotics/challenge2026_assets |
| 📊 Training dataset | datasets/ | UBTECH-Robotics/challenge2026_dataset |
Recommended Configuration
| Minimum | Recommended | Ideal | |
|---|---|---|---|
| OS | Ubuntu 22.04 / 24.04; Windows 10 / 11 | Ubuntu 22.04 / 24.04; Windows 10 / 11 | Ubuntu 22.04 / 24.04; Windows 10 / 11 |
| CPU | Intel Core i7 (7th Gen); AMD Ryzen 5 | Intel Core i7 (9th Gen); AMD Ryzen 7 | Intel Core i9 (X-series or higher); AMD Ryzen 9 / Threadripper (or higher) |
| Cores | 4 | 8 | 16 |
| RAM | 32GB | 64GB | 64GB |
| Storage | 50GB SSD | 500GB SSD | 1TB NVMe SSD |
| GPU | GeForce RTX 4080 | GeForce RTX 5080 | RTX PRO 6000 Blackwell |
| VRAM | 16GB | 16GB | 48GB |
| Driver | Linux: 580.65.06; Windows: 580.88 | Linux: 580.65.06; Windows: 580.88 | Linux: 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.
Tool Requirements
| Tool | Version | Notes |
|---|---|---|
CUDA | 12.8 | Official Guide |
Docker | latest | Official Guide |
NVIDIA Container Toolkit | latest | Official Guide |
Hugging Face | latest | pip install huggingface-hub; huggingface-cli --help (verify installation) |
Git | latest | sudo apt update; sudo apt install git -y; git --version (verify version) |
Miniconda | latest | Official Guide (optional) |
Technical Documentation Index
The complete baseline workflow consists of seven stages. We recommend following the documents in order; you may also jump directly to the stage you need based on your current progress.
| # | Document | Description |
|---|---|---|
| 1 | Resource Download | Project overview, key capabilities, hardware requirements, tool requirements, and HuggingFace resource links. |
| 2 | Environment Setup | Clone the repository, download simulation assets and datasets, build the Docker image, and configure keyboard evdev. |
| 3 | Quick Start | Start the runtime environment, keyboard teleoperation, data recording, dataset replay, and key mappings. |
| 4 | Model Training | Training guides for ACT, Diffusion Policy, π₀ (PI0), π₀.₅ (PI05), and SmolVLA with full hyperparameters. |
| 5 | Policy Inference | Run inference with a trained policy model and automatically record the results. |
| 6 | 4-Camera Real-time Display | Real-time preview from 4 RGB cameras and visualization configuration in teleoperation, recording, replay, and inference modes. |
| 7 | Automated Data Collection Tool | Collect task data through automated data collection scripts. |
Similar repos
Contributors
Languages
Python
99.7%