# GR00T-VisualSim2Real

VIRAL

Visual Sim-to-Real at Scale for
Humanoid Loco-Manipulation

DoorMan

Opening the Sim-to-Real Door for
Humanoid Pixel-to-Action Policy Transfer

## 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](https://arxiv.org/abs/2511.15200) | [Project](https://viral-humanoid.github.io/) | [Code](https://github.com/NVlabs/GR00T-VisualSim2Real/tree/main) - **DoorMan**: *Opening the Sim-to-Real Door for Humanoid Pixel-to-Action Policy Transfer* [Paper](https://arxiv.org/abs/2512.01061) | [Project](https://doorman-humanoid.github.io/) | [Code](https://github.com/NVlabs/GR00T-VisualSim2Real/tree/doorman) --- # 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](https://isaac-sim.github.io/IsaacLab/) (Isaac Sim 5.1), [TRL](https://github.com/huggingface/trl), and [Hydra](https://hydra.cc/) for configuration management. ## Table of Contents - [GR00T-VisualSim2Real](#gr00t-visualsim2real) - [Overview](#overview) - [VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation](#viral-visual-sim-to-real-at-scale-for-humanoid-loco-manipulation) - [Table of Contents](#table-of-contents) - [Prerequisites](#prerequisites) - [Installation](#installation) - [1. Create conda environment](#1-create-conda-environment) - [2. Install Isaac Sim 5.1](#2-install-isaac-sim-51) - [3. Install Isaac Lab](#3-install-isaac-lab) - [4. Install this package](#4-install-this-package) - [5. Verify installation](#5-verify-installation) - [Usage](#usage) - [Teacher Training (PPO)](#teacher-training-ppo) - [Teacher Evaluation](#teacher-evaluation) - [Student Training (DAgger Distillation)](#student-training-dagger-distillation) - [Student Evaluation](#student-evaluation) - [Configuration](#configuration) - [Experiment Tracking](#experiment-tracking) - [ONNX Export](#onnx-export) - [Project Structure](#project-structure) - [License](#license) - [Copyright \& Attribution](#copyright--attribution) - [Citation](#citation) ## Prerequisites - Ubuntu 22.04 - NVIDIA GPU with driver >= 535 - [Isaac Sim 5.1](https://docs.omniverse.nvidia.com/isaacsim/latest/installation/install_workstation.html) - [Isaac Lab](https://isaac-sim.github.io/IsaacLab/) - Conda or Mamba ## Installation ### 1. Create conda environment ```bash conda create -n viral python=3.11 -y conda activate viral ``` ### 2. Install Isaac Sim 5.1 ```bash 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](https://github.com/isaac-sim/IsaacLab) and install: ```bash pip install setuptools poetry-core flatdict pip install --no-build-isolation -e /source/isaaclab pip install --no-build-isolation -e /source/isaaclab_assets \ -e /source/isaaclab_tasks \ -e "/source/isaaclab_rl[all]" pip install numpy==1.26.0 ``` Verify the install: ```bash python -c "import isaaclab; print(isaaclab.__file__)" ``` ### 4. Install this package ```bash cd pip install -e . pip install numpy==1.26.0 # pip may upgrade it; pin again ``` ### 5. Verify installation ```bash 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: ```bash 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=False` to 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 ```bash python gr00t/rl/eval_agent_trl.py \ +checkpoint=logs_rl//model_step_044500.pt ``` ### Student Training (DAgger Distillation) Train a vision-based student policy by distilling from a trained teacher: 1. Update the teacher checkpoint path in the experiment config: ```yaml # gr00t/rl/config/exp/loco_manip/wsdpt_student_for_teacher_v8q8.002_resnet_rgb_delay.yaml teacher_actor_path: logs_rl//model_step_XXXXXX.pt ``` 2. Launch training: ```bash 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 ```bash python gr00t/rl/eval_agent_trl.py \ +checkpoint=logs_rl//model_step_XXXXXX.pt ``` ## Configuration This project uses [Hydra](https://hydra.cc/) for configuration. Configs are composed from YAML files in `gr00t/rl/config/`. Override any value from the command line: ```bash # 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](https://wandb.ai/) by default: ```bash wandb login ``` Checkpoints are saved to `logs_rl//` 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: ```bash python gr00t/rl/eval_agent_trl.py \ +checkpoint= \ num_envs=1 ``` The exported model is saved to `/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](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: ```bibtex @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} } ```