---
license: apache-2.0
tags:
- robotics
- lerobot
- robotwin
- robodojo
- action-representation
---

# Action Representation Benchmark Data

Data catalog for the `myendless/action_repre` benchmark.

## ALOHA / RoboTwin clean training data

`aloha_clean/source_robotwin_gt_depth` is a symlink in the local workspace to
`/media/damoxing/datasets/vae4d/lerobot-vae4d-org/robotwin_gt_depth`.
The benchmark uses the LeRobot Parquet export, not the raw RoboTwin HDF5 tree:
50 canonical tasks × 50 clean episodes = 2,500 episodes. Each episode provides
16D state/action targets and three calibrated camera streams. The checked source
paths are recorded in `metadata/clean50_manifest.json`.

## ARX X5 / RoboDojo data

`arx_robodojo_lerobot_v3/source_robodojo_lerobot_v3_hdf5` is a symlink in the
local workspace to the converted LeRobot v3 store at
`/media/damoxing/datasets/LVTA/datasets/robodojo_lerobot_v3_hdf5`.
It contains 3,500 episodes across 35 RoboDojo tasks, with float32 16D state and
action Parquet files plus three-camera JPEG HDF5 image shards and calibration.
This is the available ARX LeRobot-format source. It is not the same task set as
RoboTwin Core-8; the task intersection is empty.

The raw ARX RoboTwin HDF5 source is intentionally catalogued separately and is
not linked here because it is not LeRobot format.

## Reproducibility

See `metadata/dataset_catalog.json` and the benchmark manifests in the source
repositories. Local symlinks are convenience links; a Hub clone contains the
catalog and metadata, while the source data must be uploaded as regular files
or materialized by the synchronization script.