Datasets:

nvidia
/

G1-Factory-Benchmark

domain: other
modality:
  - video
  - text
  - tabular
  - timeseries
task_categories:
  - other
source: human
size_bin: n<1K
license: apache-2.0
language:
  - en
tags:
  - robotics
  - simulation
  - isaaclab-arena

G1 Factory Benchmark

Dataset Description

G1 Factory Benchmark contains 900 human-teleoperated demonstrations collected entirely in simulation using a Unitree G1 robot model, with 100 demonstrations per task across nine industrial locomanipulation tasks. The dataset provides head-camera RGB video, robot states, actions, and English task instructions in LeRobot v2.1 format.

This dataset is ready for commercial or non-commercial use.

Dataset Owner(s)

NVIDIA Corporation

Dataset Creation Date

2026-10-20 (initial upload commit on Hugging Face, UTC).

Version

v1.0

Previous Version(s): No previous version

License/Terms of Use

Apache License, Version 2.0.

Intended Usage

Training and fine-tuning robot policies, including vision-language-action models such as GR00T, for whole-body humanoid locomanipulation within the G1 Factory simulation environment only.

Dataset Characterization

Data Collection Method: Manually-Collected through human teleoperation in simulation.

Labeling Method: Hybrid: Automated, Manually-Collected

Task instructions are manually defined; simulator states and actions are recorded automatically.

Human operators used Isaac Lab-Arena, Isaac Teleop, and a Meta Quest headset. No physical robots or pre-existing seed demonstrations were used. Human operators supplied task motions, while a whole-body controller handled balance and locomotion; no generative model or autonomous task policy synthesized the demonstrations.

Dataset Format

Modality: video, text, tabular, timeseries

Format: LeRobot v2.1; Parquet, JSON, JSONL, MP4

Each task is stored at datasets/<task>/lerobot/. Episodes contain 640 × 480 H.264 RGB video at 50 Hz, 43-value joint states and targets, end-effector poses, and navigation, height, and torso commands. JSON and JSONL metadata define the features, modality mappings, and English task instructions.

Dataset Quantification

Record Count: 900 episodes, 100 per task.

Frame Count: 399,896 records at 50 Hz.

Feature Count: 17 features: 16 Parquet columns and one video stream.

Total Data Storage: Approximately 1.30 GB.

Total Video Duration: Approximately 2 hours and 13 minutes.

Task	Description
`LMBoxLift`	Lift a box from a table
`LMBoxLiftFloor`	Lift a box from the floor
`LMBoxTableToShelfPnP`	Move a box from a table to a shelf
`LMDrillLift`	Lift a drill from a table
`LMDrillLiftObs`	Navigate around a shelf and lift a drill
`LMDrillPnP`	Move a drill between tables
`LMPickDrillFromHolder`	Retrieve a drill from its holder
`LMPushButton`	Press the control-console button
`LMPushShelfForward`	Push a wheeled shelf to a target area

Usage

From the G1 Factory source repository:

git lfs install
mkdir -p datasets
git clone https://huggingface.co/datasets/nvidia/G1-Factory-Benchmark datasets/G1-Factory-Benchmark

See the G1 Factory README for fine-tuning and evaluation commands.

Reference(s)

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

This dataset is intended for robot policy training and evaluation in the G1 Factory simulation environment only. It is not intended for physical robot operation or mission-critical applications requiring functional safety. It was collected entirely in simulation and does not establish policy performance on physical hardware. Policies trained on it may not generalize to different scenes, object configurations, lighting conditions, or camera viewpoints.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.