RoboCasa GR-1 Tabletop Tasks

This repository contains the official release of simulation environments for the GR-1 Tabletop Tasks developed for NVIDIA's general-purpose humanoid foundation models: "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots." Selected task setup images are shown above.

[Website] [News] [Paper]

For business inquiries, please submit this form: NVIDIA Research Licensing

Built upon the RoboCasa simulation framework, this repository is a fork that extends its modular infrastructure to support the GR00T-N1.5-3B model. We augment RoboCasa's core functionality with additional environments, assets, and tooling tailored for the GR-1 Tabletop tasks, enabling evaluation of generalist robotic policies in diverse household tasks.

[!NOTE] The praxis branch is packaged as praxis-robocasa-gr1 and uses the robocasa_gr1 Python namespace. This keeps it independent from RoboCasa365's robocasa namespace so both benchmarks can be installed in one Praxis environment. The canonical NVIDIA / RoboCasa environment remains available on the upstream main branch.

Getting Started

RoboCasaGR00TEnvs works across all major computing platforms. The easiest way to set up is through the Anaconda package management system. Follow the instructions below to install all three required repositories, their dependencies, and download the assets needed for the simulation task:

# Install the Praxis-compatible benchmark and its simulation dependencies.
pip install "git+ssh://git@github.com/Chaoqi-LIU/robocasa-gr1-tabletop-tasks.git@praxis"

# Download the tabletop assets into the installed package.
python -m robocasa_gr1.scripts.download_tabletop_assets -y

Demo All Tasks

Run the following command to demo any available task. It sends random actions to the robot and renders an egocentric video, offering a quick, visual understanding of each task:

python3 -m robocasa_gr1.scripts.demo_task <TASK_NAME>

A list of available task names can be found later in this README.

Playback of Demonstration Trajectories

The GR00T Teleop Simulation Dataset contains an existing collection of teleoperated demonstrations spanning all 24 tabletop tasks, with 1000 human-collected demos per task.

Run the following command to play back a demo:

python3 -m robocasa_gr1.scripts.playback_dataset --dataset <HDF5_FILE> --n 1

This script replays the demonstration trajectories in simulation, rendering the robot’s actions and camera observations in sync. It’s a useful tool for visually inspecting the dataset and verifying its structure before using it for training or benchmarking.

Simulation-based Evaluation for GR00T-N1.5-3B

The main purpose of this repository is to evaluate the model in simulation to better understand its behavior in closed-loop settings. This is especially useful for assessing quantitative performance on long-horizon or multi-step tasks.

Please refer to https://github.com/NVIDIA/Isaac-GR00T to install Isaac-GR00T.

Inside the Isaac-GR00T repository, run the inference server:

cd Isaac-GR00T
python3 scripts/inference_service.py --server \
    --model_path <MODEL_PATH> \
    --data_config fourier_gr1_arms_waist

Inside the Isaac-GR00T repository, run the simulation evaluation script to evaluate a single task with 10 episodes.

cd Isaac-GR00T
python3 scripts/simulation_service.py --client \
    --env_name <TASK_NAME> \
    --video_dir ./videos \
    --max_episode_steps 720 \
    --n_envs 5 \
    --n_episodes 10

This script will run the model in a simulated environment for a given number of episodes, collect success metrics, and save rollout videos for inspection. It's a complementary method to the offline evaluation that gives insight into how the policy performs when interacting with the environment. Here is a full list of 24 tabletop task names:

gr1_unified/PnPCupToDrawerClose_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PnPPotatoToMicrowaveClose_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PnPMilkToMicrowaveClose_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PnPBottleToCabinetClose_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PnPWineToCabinetClose_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PnPCanToDrawerClose_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromCuttingboardToBasketSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromCuttingboardToCardboardboxSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromCuttingboardToPanSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromCuttingboardToPotSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromCuttingboardToTieredbasketSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromPlacematToBasketSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromPlacematToBowlSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromPlacematToPlateSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromPlacematToTieredshelfSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromPlateToBowlSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromPlateToCardboardboxSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromPlateToPanSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromPlateToPlateSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromTrayToCardboardboxSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromTrayToPlateSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromTrayToPotSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromTrayToTieredbasketSplitA_GR1ArmsAndWaistFourierHands_Env
gr1_unified/PosttrainPnPNovelFromTrayToTieredshelfSplitA_GR1ArmsAndWaistFourierHands_Env

Post-training with Isaac-GR00T

We are using the latest GR00T-N1.5-3B model from https://github.com/NVIDIA/Isaac-GR00T.git

This example finetunes a series of tasks from the Humanoid robot tabletop manipulation: 240k trajectories dataset. This runs on a H100 Node with 8 GPUs. Generally the zero-shot performance is already decent with average success rate of 42%, with this postraining it can reach higher success rate of 47%.

#!/bin/bash
ALL_DATASET_PATHS=(
  "YOUR_DATASET_PATH_1"
  "YOUR_DATASET_PATH_2"
)

python scripts/gr00t_finetune.py \
  --dataset-path "${ALL_DATASET_PATHS[@]}" \
  --num-gpus 8 --batch-size 60 --learning_rate 3e-5 \
  --output-dir OUTPUT_DIR \
  --data-config fourier_gr1_arms_waist --embodiment_tag gr1 \
  --tune-visual \ # tune visual encoder, can be omitted if you don't want to tune the visual encoder
  --max-steps 30000 --save-steps 5000

Citation

@misc{nvidia2025gr00tn1openfoundation,
      title={GR00T N1: An Open Foundation Model for Generalist Humanoid Robots}, 
      author={NVIDIA and Johan Bjorck and Fernando Castañeda and Nikita Cherniadev and Xingye Da and Runyu Ding and Linxi "Jim" Fan and Yu Fang and Dieter Fox and Fengyuan Hu and Spencer Huang and Joel Jang and Zhenyu Jiang and Jan Kautz and Kaushil Kundalia and Lawrence Lao and Zhiqi Li and Zongyu Lin and Kevin Lin and Guilin Liu and Edith Llontop and Loic Magne and Ajay Mandlekar and Avnish Narayan and Soroush Nasiriany and Scott Reed and You Liang Tan and Guanzhi Wang and Zu Wang and Jing Wang and Qi Wang and Jiannan Xiang and Yuqi Xie and Yinzhen Xu and Zhenjia Xu and Seonghyeon Ye and Zhiding Yu and Ao Zhang and Hao Zhang and Yizhou Zhao and Ruijie Zheng and Yuke Zhu},
      year={2025},
      eprint={2503.14734},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2503.14734}, 
}