GR00T-VisualSim2Real
|
VIRAL 
|
DoorMan 
|
Overview
This repository contains the application code for VIRAL (Visual Sim-to-Real for Humanoid Loco-Manipulation) and DoorMan. The system enables humanoid robots (e.g., Unitree G1) to perform complex tasks like opening heavy doors in the real world through a teacher-student simulation framework.
This repository contains the official code for:
- VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation Paper | Project | Code
- DoorMan: Opening the Sim-to-Real Door for Humanoid Pixel-to-Action Policy Transfer Paper | Project | Code
VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation
A reinforcement learning framework for humanoid robot loco-manipulation on the Unitree G1 robot. The codebase supports:
- Teacher Training -- PPO-based policy training with privileged state observations
- Student Training -- Vision-based policy distillation (DAgger) from a trained teacher using RGB camera input
- Evaluation -- Evaluate trained teacher or student checkpoints, with optional ONNX export for deployment
Built on Isaac Lab (Isaac Sim 5.1), TRL, and Hydra for configuration management.
Table of Contents
- GR00T-VisualSim2Real
- VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation
Prerequisites
- Ubuntu 22.04
- NVIDIA GPU with driver >= 535
- Isaac Sim 5.1
- Isaac Lab
- Conda or Mamba
Installation
1. Create conda environment
conda create -n viral python=3.11 -y
conda activate viral
2. Install Isaac Sim 5.1
pip install torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu128
pip install isaacsim==5.1.0.0 isaacsim-rl==5.1.0.0
3. Install Isaac Lab
Clone or download Isaac Lab and install:
pip install setuptools poetry-core flatdict
pip install --no-build-isolation -e <path-to-IsaacLab>/source/isaaclab
pip install --no-build-isolation -e <path-to-IsaacLab>/source/isaaclab_assets \
-e <path-to-IsaacLab>/source/isaaclab_tasks \
-e "<path-to-IsaacLab>/source/isaaclab_rl[all]"
pip install numpy==1.26.0
Verify the install:
python -c "import isaaclab; print(isaaclab.__file__)"
4. Install this package
cd <path-to-this-repo>
pip install -e .
pip install numpy==1.26.0 # pip may upgrade it; pin again
5. Verify installation
python -c "from gr00t.rl.envs.base_task.base_task import BaseTask; print('OK')"
Usage
Teacher Training (PPO)
Train a teacher policy using privileged state observations:
HYDRA_FULL_ERROR=1 accelerate launch --num_processes 1 \
gr00t/rl/train_agent_trl.py \
+exp=loco_manip/walk_stand_place_grasp_turn_homie \
num_envs=48 \
project_name=wsdpt_teacher
Tip: Add
headless=Falseto open the Isaac Sim GUI and watch training live.
Teacher policy running in Isaac Sim
| Argument | Description |
|---|---|
num_envs |
Number of parallel environments (higher = faster, more VRAM) |
project_name |
Weights & Biases project name |
headless |
True (default) for headless; False to open GUI |
env.config.reset_from_dataset.enable |
Reset from demonstration dataset |
Teacher Evaluation
python gr00t/rl/eval_agent_trl.py \
+checkpoint=logs_rl/<experiment_dir>/model_step_044500.pt
Student Training (DAgger Distillation)
Train a vision-based student policy by distilling from a trained teacher:
- Update the teacher checkpoint path in the experiment config:
# gr00t/rl/config/exp/loco_manip/wsdpt_student_for_teacher_v8q8.002_resnet_rgb_delay.yaml
teacher_actor_path: logs_rl/<your_teacher_experiment>/model_step_XXXXXX.pt
- Launch training:
HYDRA_FULL_ERROR=1 accelerate launch --num_processes 1 \
gr00t/rl/train_agent_trl.py \
+exp=loco_manip/wsdpt_student_for_teacher_v8q8.002_resnet_rgb_delay \
num_envs=8 \
headless=True \
experiment_name=wsdpt_student \
project_name=wsdpt_student_debug
If you add headless=False, you can see the student policy running in Isaac Sim:
Student policy running in Isaac Sim
Student Evaluation
python gr00t/rl/eval_agent_trl.py \
+checkpoint=logs_rl/<student_experiment_dir>/model_step_XXXXXX.pt
Configuration
This project uses Hydra for configuration. Configs are composed from YAML files in gr00t/rl/config/. Override any value from the command line:
# Number of environments
... num_envs=16
# Reward weights
... rewards.reward_scales.tracking_lin_vel=1.0
# Training hyperparameters
... algo.config.actor_learning_rate=1e-4
Experiment Tracking
Training logs to Weights & Biases by default:
wandb login
Checkpoints are saved to logs_rl/<experiment_name>/ at intervals controlled by the save_frequency callback parameter.
ONNX Export
During evaluation with num_envs=1, the policy is automatically exported as ONNX for deployment:
python gr00t/rl/eval_agent_trl.py \
+checkpoint=<path_to_checkpoint.pt> \
num_envs=1
The exported model is saved to <experiment_dir>/exported/.
Project Structure
gr00t/rl/
├── train_agent_trl.py # Training entry point (teacher & student)
├── eval_agent_trl.py # Evaluation entry point
├── config/ # Hydra YAML configs
│ ├── base.yaml # Base training config
│ ├── base_eval.yaml # Base evaluation config
│ ├── exp/loco_manip/ # Experiment configs
│ ├── algo/ # Algorithm configs (PPO, DAgger)
│ ├── env/ # Environment configs
│ ├── robot/g1/ # Robot configs (G1 43-DOF)
│ ├── rewards/ # Reward function configs
│ ├── terrain/ # Terrain configs
│ ├── obs/ # Observation configs
│ └── domain_rand/ # Domain randomization configs
├── envs/ # Environment implementations
│ ├── base_task/ # Base task classes
│ └── loco_manip/ # Loco-manipulation task
├── trl/ # TRL-based trainers and modules
│ ├── trainer/ # PPO and distillation trainers
│ ├── modules/ # Actor-critic network modules
│ ├── callbacks/ # Training callbacks
│ └── utils/ # Training utilities
├── agents/modules/ # Neural network building blocks
├── simulator/isaacsim/ # Isaac Sim interface
├── data/ # Task data (robot assets, scenarios)
└── utils/ # General utilities
License
This project is licensed under the Apache License 2.0. See the LICENSE file for details.
Copyright & Attribution
Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
This project includes third-party open-source software. Please refer to individual source files for specific licenses and copyright headers.
Citation
If you find this work useful, please cite:
@inproceedings{he2025viral,
title={VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation},
author={He, Tairan and Wang, Zi and Xue, Haoru and Ben, Qingwei and Luo, Zhengyi and Xiao, Wenli and Yuan, Ye and Da, Xingye and Castañeda, Fernando and Sastry, Shankar and Liu, Changliu and Shi, Guanya and Fan, Linxi and Zhu, Yuke},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2026}
}
@inproceedings{xue2025opening,
title={Opening the Sim-to-Real Door for Humanoid Pixel-to-Action Policy Transfer},
author={Xue, Haoru and He, Tairan and Wang, Zi and Ben, Qingwei and Xiao, Wenli and Luo, Zhengyi and Da, Xingye and Casta{\~n}eda, Fernando and Shi, Guanya and Sastry, Shankar and others},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2026}
}