---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- yam
- bimanual
- rubiks-cube
- curated
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# Magic Cube

YAM bimanual demonstrations: pick up a Rubik's cube with the left arm, present it to the right arm, and turn an outer layer with the right gripper.

This is the stationary-pause-filtered, 2x temporal-subsampling edition of [zekaili/yam-cube-layer-turn](https://huggingface.co/datasets/zekaili/yam-cube-layer-turn), based on source revision `e0afe40d7f2402424c9920e5a49e381de8816dfb`.

| Property | Value |
| --- | --- |
| Format | LeRobot v3.0 |
| Episodes | 46 |
| Frames | 33,166 |
| FPS | 15 |
| Total duration | 2,211.067 seconds (36 min 51 sec) |
| Views | top, left_wrist, right_wrist |
| State/action | 12 arm joints plus left/right grippers, original recorded values |
| Video | H.264, lossless relative to decoded retained source frames |

## Processing

One original episode remains one output episode. Long stationary pause interiors are removed synchronously from all videos and numeric rows, then retained intervals are concatenated in source order. Every other frame of that filtered episode is selected, starting at frame zero, while output FPS remains 15. State/action values are selected, not multiplied or interpolated. Frame indices, global indices, timestamps, and statistics are rebuilt.

The filter requires both arms and grippers to remain stationary for at least 3 seconds, with approximately 0.5 seconds of context retained at each edge. It checks measured state motion and action-target changes, including whole-pause drift limits. Detailed native-unit thresholds are in `meta/curation/plan.json`. Defaults were calibrated on source episode 0; the full dataset was processed automatically, not exhaustively visually reviewed.

The original 165,746 numeric frames became 66,308 after pause removal and common-prefix handling, then 33,166 after 2x subsampling. There were 99,435 pause frames removed, 3 additional alignment-tail rows removed, and 33,142 frames skipped by subsampling. Integer rounding is applied independently to each episode; an extra endpoint frame is not appended.

Temporal compression provides faster demonstrations at the dataset timebase. Actual robot execution speed depends on the controller timing and limits; this export has not been validated by a physical robot execution test.

## Source alignment caveat

Source episode 17/top, 20/right_wrist, and 40/right_wrist (zero-based indices) each contain one fewer video frame than numeric rows. Those episodes use the common prefix across views and numeric rows; the last numeric row and corresponding longer-view frames are omitted before filtering. The original missing capture position is unknown, so these three episodes inherit up to one source frame (about 67 ms) of alignment uncertainty. This is recorded in the per-episode plan and verification files.

## Verification and provenance

All 138 exported videos were checked frame by frame in decoded YUV and RGB against the selected source video frames. Retained state/action/task values and source-frame mappings were independently checked, and episode 0 exactly reproduces the earlier 2x pilot. The original dataset was preserved. These checks verify the transformation, not the original sensor alignment.

- `meta/curation/plan.json`: parameters, pause selections, subsampling, joins, and alignment handling.
- `meta/curation/source_frame_mapping.parquet`: every output frame's original episode/frame/global index and original timestamp, plus its position on the pause-filtered timeline.
- `meta/curation/video_verification.json`: decoded pixel hashes and video verification.
- `meta/curation/verification.json`: export counts and checks.
- `meta/curation/source_manifest.json`: original dataset file hashes.
- `meta/curation/code/filter_lerobot_pauses.py`: processing code snapshot; run within the Sana project with its referenced helpers.

Numeric statistics are recomputed from retained rows. Image statistics use stride-4 pixel sampling of retained frames. Audit records retain original storage paths as provenance; LeRobot training paths in `meta/info.json` are relative and portable.

Apache-2.0 is inherited from the source dataset card.