README.md

---
license: apache-2.0
pretty_name: Anchor-Lab
size_categories:
- "100K<n<1M"
tags:
- robotics
- sim-to-real
- h1
- g1
- unitree
- RL
- isaaclab
- newton
- system-identification
- robot-calibration
- physical-ai
configs:
- config_name: default
  data_files:
  - split: train
    path: data/**/*.parquet
---

# Anchor-Lab Dataset Card

## Dataset Description

Anchor-Lab is a tabular robotics dataset captured from Anchor Lab, a sim-to-real transfer laboratory from NVIDIA Robotics. The dataset is designed to support calibration of physics simulation against physical robot measurements for zero-shot sim-to-real deployment.

The public release contains long-form parquet tables of robot experiment telemetry. Each row records a timestamped scalar measurement for an experiment and signal field. The dataset covers H1- and G1-family humanoid robot experiments, including arm, leg, multi-joint, single-joint, and elbow test-stand measurements.

For simplicity, this card refers to the dataset as Anchor-Lab.

## Dataset Group Overview

The release is organized into seven top-level data groups under `data/`.

| Group | Description | Approx. storage |
|---|---|---:|
| `g1_arm_multijoint` | G1 arm multi-joint excitation trials, including multi-frequency, multisine, chirp, wave-sine, wave-chirp, antiphase, and related patterns. | 53.3 MB |
| `g1_arm_single_joint` | G1 arm single-joint trials across elbow, shoulder pitch, shoulder roll, and related arm joints with pose and excitation variations. | 238 MB |
| `g1_leg_multijoint` | G1 leg multi-joint trials across pose variants, including independent multisine, multi-frequency, slow-wave, wave-sine, and wave-chirp patterns. | 85.9 MB |
| `g1_leg_single_joint` | G1 leg single-joint trials across ankle, hip, knee, and related leg joints with pose and excitation variations. | 479 MB |
| `h1_elbow_teststand` | H1 elbow test-stand system-identification measurements across frequency, amplitude, and position settings. | 334 MB |
| `h1_multijoint` | H1 multi-joint trials with chirp and sine-style excitations across frequency, amplitude, in-phase, antiphase, and wave settings. | 318 MB |
| `h1_single_joint` | H1 single-joint trials for elbow, shoulder pitch, shoulder raise, and related joints using sine and chirp-style excitations. | 349 MB |


## Dataset Category Previews

<table>
  <tr>
    <td align="center"><strong>G1 Dataset</strong></td>
    <td align="center"><strong>H1 Dataset</strong></td>
    <td align="center"><strong>Isolated Motor / Test-Stand Dataset</strong></td>
  </tr>
  <tr>
    <td><img src="./G1.gif" alt="G1 sim-to-real recording preview" width="100%"></td>
    <td><img src="./H1.gif" alt="H1 sim-to-real recording preview" width="100%"></td>
    <td><img src="./isolated_motor.gif" alt="Isolated motor and test-stand recording preview" width="100%"></td>
  </tr>
  <tr>
    <td>G1 dataset (G1 arm single-joint and multi-joint).</td>
    <td>H1 dataset (H1 elbow single-joint and multi-joint).</td>
    <td>Benchtop isolated H1 motor dataset (test-stand).</td>
  </tr>
</table>

These previews are included for qualitative inspection of the experimental setup and motion patterns. They are not required for loading, filtering, or analyzing the parquet data.

## Dataset Owner(s)

NVIDIA Corporation / NVIDIA Robotics

## Dataset Creation Date

Public release: June 2026

## Version

v1.0

Previous version(s): N/A

## License/Terms of Use

Apache License 2.0

## Intended Usage

This dataset is intended for researchers and AI practitioners working on:

- Simulation calibration against measured physical robot behavior.
- System identification for humanoid robot joints, actuators, and coupled multi-joint dynamics.
- Sim-to-real transfer workflows for zero-shot or few-shot physical deployment.
- Validation of physics-engine, actuator, controller, and sensor models against laboratory measurements.
- Robot-control research using timestamped measurements from H1 and G1 humanoid platforms.
- Analysis pipelines for long-form robot telemetry in parquet format.

This dataset is not intended to be used as the sole validation source for safety-critical robot deployment. Controllers, policies, and calibrated models derived from this dataset should be evaluated in simulation, test stands, and controlled physical environments before any real-world deployment.

## Dataset Characterization

### Data Collection Method

Physical - The data was captured from Anchor Lab experiments using onboard sensors found in the Unitree platforms and test-stand measurement pipelines. The experiment identifiers and file names encode platform, body region, single-joint or multi-joint scope, joint or joint group, excitation type, pose variant, trial number, and, where available, collection date, frequency, amplitude, or position setting.

The public data is organized around structured excitation and system-identification-style trials. File names indicate patterns such as sine, chirp, multisine, sine sweep, etc.

### Labeling Method

Automated / procedural - The released labels are metadata generated by the acquisition and data-processing pipeline. The core metadata fields are:

- `experiment`: trial or experiment identifier.
- `field`: measured signal name.
- `time_ns` and `time_utc`: timestamp fields.
- `value`: scalar measurement value for the given timestamp and field.

The dataset does not provide human-annotated semantic labels. Field names and experiment identifiers should be treated as structured telemetry metadata.

## Dataset Category Description

### G1 Arm - Multi-Joint

The `g1_arm_multijoint` group contains multi-joint arm excitation trials for G1. Visible file naming patterns include antiphase, broadband pose, fast wave, multi-chirp, multi-frequency, multisine, slow wave, wave-chirp, and wave-sine trials.

### G1 Arm - Single Joint

The `g1_arm_single_joint` group contains single-joint G1 arm measurements across elbow, shoulder pitch, shoulder roll, and related arm joints. Visible file naming patterns include pose variants such as neutral, forward-raised-extended, and back-bent, and excitation types such as bookend, chirp, multisine, sine sweep, slow ramp, step, and varying amplitude.

### G1 Leg - Multi-Joint

The `g1_leg_multijoint` group contains multi-joint G1 leg excitation trials. Visible file naming patterns include independent multisine, multi-frequency, slow-wave, wave-chirp, and wave-sine experiments across pose variants.

### G1 Leg - Single Joint

The `g1_leg_single_joint` group contains single-joint G1 leg measurements across ankle, hip, knee, and related leg joints. Visible file naming patterns include ankle pitch, ankle roll, hip pitch, and pose variants such as neutral, forward-bent, and back-extended.

### H1 Elbow Test Stand

The `h1_elbow_teststand` group contains H1 elbow test-stand system-identification measurements. File names encode settings such as frequency, amplitude level, and position index.

### H1 Multi-Joint

The `h1_multijoint` group contains H1 multi-joint experiments. Visible file naming patterns include chirp and sine trials with frequency ranges, amplitude values, and phase/coordination variants such as in-phase, wave, and antiphase.

### H1 Single Joint

The `h1_single_joint` group contains H1 single-joint experiments for elbow, shoulder pitch, shoulder raise, and related joints. Visible file naming patterns include chirp and sine trials across pose variants and repeated trial indices.

## Dataset Format

| Field | Type | Description |
|---|---|---|
| `time_ns` | `int64` | Nanosecond-resolution timestamp for the measurement row. |
| `time_utc` | `timestamp[ns, tz=UTC]` | UTC timestamp corresponding to `time_ns`. |
| `experiment` | `string` | Experiment or trial identifier. |
| `field` | `string` | Name of the measured signal. |
| `value` | `float64` | Scalar measurement value for the given timestamp, experiment, and field. |

The dataset is stored in long format: one row corresponds to one scalar value for one signal field at one timestamp in one experiment. Users who need a wide time-series table can pivot by `field` after loading a subset.

## Dataset Organization

```text
Anchor-Lab/
├── README.md
├── G1.gif
├── H1.gif
├── isolated_motor.gif
└── data/
    ├── g1_arm_multijoint/
    ├── g1_arm_single_joint/
    ├── g1_leg_multijoint/
    ├── g1_leg_single_joint/
    ├── h1_elbow_teststand/
    ├── h1_multijoint/
    └── h1_single_joint/
```

Each leaf directory contains parquet files for individual experiments or trials. File names generally follow this pattern, with group-specific variations:

```text
<robot>-<body-region>-<single-or-multijoint>-<joint-or-excitation>-<pose-or-condition>-<trial-or-settings>.parquet
```

Examples:

```text
data/h1_multijoint/h1-multijoint-v1-chirp-trial-001-freq-0p1-to-0p5hz-amp-0p05.parquet
data/g1_leg_multijoint/g1-leg-multijoint-wave-sine-pose-a-trial-001-20260408-173950.parquet
data/g1_arm_single_joint/g1-arm-single-joint-elbow-step-pose-a-neutral-20260403-091255.parquet
```

## Download And Loading Examples

Install optional dependencies:

```bash
pip install -U "huggingface_hub[cli]" datasets pandas pyarrow polars
```

Download the complete dataset:

```bash
hf download nvidia/Anchor-Lab \
  --repo-type dataset \
  --local-dir ./Anchor-Lab
```

Download one source group, for example H1 multi-joint data:

```bash
hf download nvidia/Anchor-Lab \
  --repo-type dataset \
  --include "README.md" "data/h1_multijoint/*" \
  --local-dir ./Anchor-Lab-h1-multijoint
```

Other useful subset patterns include:

```text
data/g1_arm_multijoint/*
data/g1_arm_single_joint/*
data/g1_leg_multijoint/*
data/g1_leg_single_joint/*
data/h1_elbow_teststand/*
data/h1_multijoint/*
data/h1_single_joint/*
```

Download a subset from Python:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="nvidia/Anchor-Lab",
    repo_type="dataset",
    local_dir="./Anchor-Lab-g1-leg-single-joint",
    allow_patterns=[
        "README.md",
        "data/g1_leg_single_joint/*.parquet",
    ],
)
```

Load through Hugging Face Datasets:

```python
from datasets import load_dataset

ds = load_dataset("nvidia/Anchor-Lab", split="train")
print(ds)
print(ds.features)
print(ds[0])
```

## Dataset Quantification

| Quantity | Count / Size |
|---|---:|
| Rows in default train split | 992,000 |
| Top-level data groups | 7 |
| Public file format | Parquet |
| Total repository file size | 1.86 GB |
| Modalities | Tabular telemetry with string metadata fields |

Approximate storage by top-level data group:

| Group | Approx. storage |
|---|---:|
| `g1_arm_multijoint` | 53.3 MB |
| `g1_arm_single_joint` | 238 MB |
| `g1_leg_multijoint` | 85.9 MB |
| `g1_leg_single_joint` | 479 MB |
| `h1_elbow_teststand` | 334 MB |
| `h1_multijoint` | 318 MB |
| `h1_single_joint` | 349 MB |
| Total | 1.86 GB |

## Curation Notes

- The dataset uses a compact long-form schema shared across the public parquet files.
- The default Hugging Face viewer exposes a single `train` split.
- Experiment identifiers and file names are important metadata and should be preserved during downstream processing.
- Users should validate timestamp monotonicity, duplicate rows, missing fields, and field coverage before fitting calibration or system-identification models.
- Units are field-specific. Do not assume a single unit convention across all `field` values without a data dictionary or experiment-specific documentation.

## Known Limitations and Future Work

- The current public schema is minimal and does not include a full data dictionary for every `field` value.
- The release does not define standardized train/validation/test splits for model benchmarking.
- Measurements may be specific to the hardware revision, firmware, controller configuration, sensors, test stand, and laboratory setup used during capture.
- Timing, filtering, actuator latency, sensor latency, and controller-loop details may affect downstream system-identification results.
- The dataset focuses on H1 and G1 humanoid robot experiments and should not be treated as representative of all robots or all operating conditions.
- Future releases could add per-file manifests, units, calibration parameters, robot configuration metadata, controller settings, sensor descriptions, and benchmark splits.

## Ethical Considerations

This dataset contains robot experiment telemetry and does not intentionally include human-subject data, real-world imagery, or personally identifiable information. However, robotics datasets can still be misused if they are used to calibrate, validate, or deploy unsafe physical controllers without adequate testing.

Users should evaluate derived models in simulation, isolated test stands, and controlled physical settings before deployment. Users are responsible for ensuring that their use of this dataset complies with applicable safety, security, privacy, and robotics deployment requirements.

## Contributors

- Elena Shrestha
- Vignesh Bhavananthan
- Vidur Vij
- Alex Omar
- Andrew Wrenn
- Justin Shrake
- Michael Clive
- Hong Wang

## Citation

Please cite the dataset as:

```bibtex
@dataset{nvidia_anchor_lab_2026,
  title        = {Anchor-Lab},
  author       = {NVIDIA},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/nvidia/Anchor-Lab}
}
```