---
pretty_name: SO101 Stationery Tidying DAgger - 2026-09-09
tags:
- lerobot
- robotics
- so101
- dagger
- imitation-learning
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*/*.parquet
---

# SO101 Stationery Tidying DAgger

65 episodes collected on two dual-SO101 setups (4090B and 4090C), merged from 14 datasets in native LeRobot v3 format.

Task: **Put the stationery on the table into the bag.**

Original task text is retained. One source uses `put stationery table into bag`; the others use `Put the stationery on the table into the bag.` Task indices are remapped to this merged task table.

| Property | Value |
|---|---|
| Collection window | 2026-09-08 22:00 to 2026-09-09 02:48:40, Asia/Shanghai |
| Episodes | 65 |
| Frames | 345,892 |
| FPS | 30 |
| Saved duration | 03:12:09.7 (11,529.733 seconds) |
| Human intervention frames | 109,466 |
| Human intervention duration | 01:00:48.9 (3,648.867 seconds) |
| Robot type | bi_so_follower |
| Action / state dimensions | 12 / 12 |
| Cameras | left_wrist, environment_1, right_wrist |
| Collection method | Episodic DAgger with policy rollout and human takeover |

The boolean `intervention` field is true for saved human-control frames and false for policy-control frames. Durations are saved frame counts divided by 30 FPS, excluding reset, waiting, and unsaved recordings.

## Sources and processing

4090B contributes 37 episodes; 4090C contributes 28 episodes. Per-episode source mappings are in `meta/source_episodes.json`, and per-dataset totals are in `meta/source_datasets.json`.

Three previously identified short episodes (23.2 s, 16.4 s, 11.0 s) were removed before this export. A 32.7-second episode from the C dataset ending in `20260909_022707` is retained; task completion has not been visually verified.

Merged using LeRobot's native aggregation utility. Video files were copied without re-encoding. Episode and global frame indices were updated, and task indices were remapped while preserving the exact task text. All other saved data columns were compared against their source values and matched. Source-to-output video SHA-256 hashes matched. Per-episode frame counts, saved timestamps, and video metadata were validated. These structural checks do not establish task success or prove that capture never paused or disconnected.