---
pretty_name: ReFrame Handheld Demonstration Dataset
language:
- en
- zh
tags:
- robotics
- imitation-learning
- video
- timeseries
- multi-view
---
# ReFrame Handheld Demonstration Dataset

ReFrame-Handheld provides real handheld operation recordings for studying how demonstrations can be reused across camera configurations. It contains **1,651 independent recordings across four manipulation tasks**, with camera images, per-frame timestamps, raw SLAM poses, gripper readings and camera calibration.

The data supports inspecting viewpoint changes, testing geometric reprojection and comparing transformed images with available real reference-camera observations. Videos and structured numerical files can be read without a ROS environment.

**1,651 recordings · 4 tasks · approximately 303 recorded minutes · 67 GB of episode data**

## Observations

- **1,451 recordings** contain main RGB, left and right grayscale cameras, and a real USB reference-camera stream.
- **200 recordings** contain main RGB without auxiliary images.
- All recordings retain SLAM pose and gripper samples. Each episode references its applicable calibration profile.

The recordings use the same brand and model of capture device with **three camera-extrinsic configurations**. The [dataset overview](docs/DATASET_OVERVIEW.md) describes the observation types, statistics and counting rules.

## Tasks

| Task | Independent recordings | Approximate episode download size |
| --- | ---: | ---: |
| Carrot picking and placing | 696 | 34.49 GB |
| Button pressing | 500 | 19.69 GB |
| Drawer opening | 255 | 9.01 GB |
| Cup hanging | 200 | 4.05 GB |
| **Total** | **1,651** | **67.25 GB** |

Sizes refer to episode files, excluding the optional starter ZIP and shared documentation. Recorded duration includes recording margins and is counted once per operation, not once per camera.

## Download and try

**[Start with 17 complete examples](downloads/reframe_handheld_starter_17.zip)** — approximately 788 MB, covering all retained acquisition batches and all four tasks. The ZIP includes calibration, reading tools and instructions, and works independently after extraction.

All episode data is distributed as TAR shards, grouped by task and acquisition batch. Each shard contains complete episodes with their original files and paths. The [download tool](docs/QUICKSTART.md) downloads the required shards, verifies SHA256 and extracts the selected episodes with their calibration. Selecting one episode downloads its containing shard; see [the archive index](metadata/shards.jsonl) for exact sizes.

Open an MP4 directly to inspect a recording. For Python access to images, timestamps, poses, gripper readings and calibration, follow the [quickstart](docs/QUICKSTART.md).

## Documentation

- [Dataset overview and statistics](docs/DATASET_OVERVIEW.md)
- [File format and reading interfaces](docs/FORMAT.md)
- [Timestamp semantics and optional frame matching](docs/TIME_ALIGNMENT.md)
- [Camera calibration](docs/CALIBRATION.md)
- [Selection, task-outcome labels and known limitations](docs/QUALITY.md)

The dataset preserves original operation sequences and quality annotations. See the linked documentation for timestamp interpretation, outcome labels and calibration applicability.

## License and citation

[TO FILL: confirmed dataset license]

[TO FILL: authors and paper citation]

[TO FILL: paper, code repository and contact]

This dataset accompanies ReFrame: Reusing Handheld Demonstrations Across Wrist-Camera Configurations. Licensing and paper citation details are pending owner confirmation.


## Additional calibration candidates

Brown5 and fisheye4 intrinsic refits are available for the Alice and David USB cameras. They are optional candidates; episode references continue to use the historical calibration profiles. See [the 2026-10-29 calibration update](docs/USB_INTRINSIC_REFITS_20261029.md).