adityx23/icl-demo-dataset

ICL Demo Dataset

A small, recent collection of bimanual manipulation demonstrations on the YOR robot, in LeRobot v2.1 format. Recorded 3 and 6 August 2026.

285 episodes · 254,171 frames · 2.35 hours · 27 tasks · 3 camera views

This is a companion to adityx23/icl-dataset — same robot, same schema, same task vocabulary, but a separate and much smaller collection. Roughly ten demonstrations per task, several of them tasks that do not appear in the larger dataset at all.

---

Outcomes

success	valid	keep	episodes	frames	what it is
✅	✅	✅	269	238,681	the training set
❌	✅	❌	15	15,216	genuine attempt, goal not achieved
❌	❌	❌	1	274	not a usable demonstration

Every episode carries three boolean flags — in meta/episodes.jsonl and as columns on every parquet row. keep is the training filter.

from datasets import load_dataset

ds = load_dataset("adityx23/icl-demo-dataset", split="train")
train = ds.filter(lambda r: r["keep"])      # 269 episodes

This dataset is not balanced, so no episode has success=true, keep=false — every success collected is kept. The flags otherwise match adityx23/icl-dataset.

---|---:|---:| | train | 269 | 238,681 | | fail | 15 | 15,216 | | invalid | 1 | 274 |

Failures and unusable episodes sit in the same files as the training set, distinguished only by the label field — present both in meta/episodes.jsonl and as a column on every parquet row. A naive load returns all 285. To train on the usable ones:

from datasets import load_dataset

ds = load_dataset("adityx23/icl-demo-dataset", split="train")
train = ds.filter(lambda r: r["label"] == "train")     # 269 episodes

This dataset is not balanced, so it has no trimmed episodes — every success collected is labelled train. The vocabulary otherwise matches adityx23/icl-dataset.

At 94% success this set is cleaner than the larger dataset, but it is also small — most tasks have only ~11 episodes, so treat it as demonstration data or a fine-tuning / evaluation set rather than a training corpus on its own.

---

Tasks

id	task (instruction)	episodes	minutes	success	fail	invalid
3	open the gatorade bottle	23	10.8	15	8	0
2	open the doctor pepper bottle	15	9.2	12	3	0
16	put the green pepper into the grocery bag	12	3.3	11	1	0
5	open the notebook and place the cube on it	12	4.8	11	0	1
7	pass the mustard to the plate	12	4.0	11	1	0
17	put the mustard into the box	12	1.6	12	0	0
23	stack the three cubes into a tower	11	6.5	11	0	0
24	stack the three cups into a tower	11	12.5	11	0	0
13	put the eggplant into the bag	11	3.3	11	0	0
12	put the circle on the peg	11	1.8	11	0	0
21	stack the colored octagons	11	8.4	11	0	0
0	hit the eggplant with the mallet	11	3.0	11	0	0
1	hit the yellow cube with the mallet using the left arm	11	4.2	11	0	0
20	sort the items into their containers	11	19.6	11	0	0
4	open the notebook	11	2.9	11	0	0
6	pass the gusset to the plate	11	3.0	11	0	0
19	sort the food items into the cups	11	11.0	10	1	0
18	put the square on the peg	11	3.8	11	0	0
22	stack the cubes and hit them with the mallet	11	10.7	11	0	0
25	uncap the black marker	11	4.1	10	1	0
26	uncap the red marker	11	5.5	11	0	0
8	pass the salt shaker from the left arm to the right arm	6	2.1	6	0	0
9	pass the salt shaker from the right arm to the left arm	6	1.9	6	0	0
15	put the eggplant into the box with the right arm	6	1.0	6	0	0
11	pick up the red chilli and place it on the plate with the right arm	6	0.7	6	0	0
14	put the eggplant into the box with the left arm	5	0.9	5	0	0
10	pick up the red chilli and place it on the plate with the left arm	5	0.6	5	0	0

Task labels are natural-language instructions, one instruction per task and one task_index per instruction, using the same vocabulary as adityx23/icl-dataset so shared tasks mean the same thing in both.

Three of the source runs alternated arm or hand-off direction episode-by-episode. Those were verified from gripper timing and split into per-episode tasks rather than being labelled with a single instruction for the run — so, for example, the salt shaker run contributes six left-to-right episodes and six right-to-left.

---

Structure

meta/
  info.json             LeRobot v2.1 header
  episodes.jsonl        one row per episode: index, task, length, outcome, provenance
  episodes_stats.jsonl  per-episode feature statistics
  tasks.jsonl           27 task instructions, indexed
data/chunk-XXX/episode_NNNNNN.parquet
videos/chunk-XXX/observation.images.<cam>/episode_NNNNNN.mp4   h264, 640x480, 30 fps

Cameras are zed (scene), fish0 (left gripper), fish1 (right gripper).

Per-frame features

feature	shape	description
observation.state	[15]	measured proprioception: lj0..lj6, rj0..rj6, lift
action.left_ee	[7]	commanded left end-effector pose, [qw,qx,qy,qz,x,y,z]
action.right_ee	[7]	commanded right end-effector pose
action.gripper	[2]	commanded [left, right], 0.0 = closed, 1.0 = open
action.base_vel	[3]	commanded base velocity [vx, vy, omega]
action.lift_cmd	[1]	commanded lift, -1 / 0 / 1
3 × observation.images.*	480x640x3	h264 video

Plus timestamp, frame_index, episode_index, index, task_index.

observation.state is measured; everything under action.* is the operator's commanded target, so gripper state lives in the action space.

Stable episode identity

episode_index is required by LeRobot to run contiguously from 0, so it necessarily changes whenever episodes are added, dropped or reordered — it has shifted meaning in this dataset's history, silently invalidating cached copies.

Every episode therefore also carries episode_uid, which does not change across rebuilds:

"episode_uid": "clean_the_plate_20260730_182850:000069"
                └─ collection run ──────────────┘ └ index within that run

It appears in two places: as a field on every record in meta/episodes.jsonl, and as an episode_uid string column in every parquet (constant within a file), so it travels with the frames themselves and is visible in the dataset viewer. That column is an addition to the standard LeRobot v2.1 schema; it is declared in meta/info.json under features, and loaders that select columns explicitly are unaffected.

Use episode_uid as the join key for anything you want to survive a dataset update — annotations, evaluation results, hand-picked subsets. Use episode_index only to locate the parquet and video files in this revision.

The uid is derived from source_dataset and source_episode, so it can be regenerated at any time rather than being an opaque stored value. It changes only if a source collection run is itself renamed or restructured.

---

Provenance

Merged from 24 collection runs. Every episode records where it came from:

Every episode record is five fields — the collection run is the first half of its episode_uid:

{"episode_index": 0, "episode_uid": "3_cube_tower_stack_20260803_114230:000000", "tasks": ["stack the three cubes into a tower"], "length": 1290, "success": true, "valid": true, "keep": true}

Episodes that were themselves folded in from a second run of the same task carry an additional source_run / source_episode pair from that merge.

---

Known limitations

• Small. ~11 episodes per task; two tasks have more only because duplicate runs of the same task were merged.
• Labels are the operator's, recorded at collection time and not independently audited. An automated screen for the failure mode that plagued the larger dataset (episodes marked success where the robot never moved) flagged one episode here, put the mustard into the box — the object is already in the box at frame 1.
• One episode carries merged_from_unknown, meaning no outcome was recorded at collection and it was folded into invalid.
• No train/validation split is defined.
• Camera extrinsics/intrinsics are not published; the two gripper fisheyes are identical models distinguished only by which USB port they occupy.

License

Not specified. No licence has been chosen for this data, so downstream use is legally ambiguous. Check with the authors before using it.

---

Revision history

2026-08-20 (later) — one label string replaced by three booleans. label is gone, split into success, valid and keep, matching adityx23/icl-dataset. keep == true selects exactly what label == "train" did. Note success changed type from string to boolean.

2026-08-20 — episode record simplified to five fields. source_dataset, source_episode, subset and balance were dropped as redundant, and the outcome field was renamed success -> label with values train / fail / invalid, matching adityx23/icl-dataset. Breaking change to field names only — counts, indices, videos and frame data are unchanged.

2026-08-19 — outcome columns added. success and balance are now columns on every parquet row (declared in meta/info.json), matching adityx23/icl-dataset, so the outcome is visible to the dataset viewer and to load_dataset() rather than living only in meta/episodes.jsonl. No episode index, task index or video changed.

2026-08-07 — episode_uid and subset added. Two string columns were added to meta/episodes.jsonl and to every parquet, and declared in meta/info.json under features (14 features -> 16), matching the schema of adityx23/icl-dataset. episode_uid gives each episode an identity that survives rebuilds; subset is empty throughout this dataset. Both are additions to the standard LeRobot v2.1 schema. No episode index, task index or video changed.