isYes/mycobot280_amr_pickplace

mycobot280 — AMR Pick-and-Place with Factored Shape / Pose / Viewpoint Variation

A real-robot imitation-learning dataset of a 7-DoF myCobot 280 picking up a casting and placing it into a bin carried by a small AMR (autonomous mobile robot).

What makes this dataset different from a typical single-task pick-and-place set is that the variation is factored and fully labelled. Every episode carries the ground-truth value of four independent scene factors — item shape, item pose, AMR position, AMR heading angle — so the dataset can be sliced along any one of them to build controlled in-distribution / out-of-distribution splits. It was collected to study whether a behaviour-cloning policy can be made robust to nuisance variation by conditioning on factors it can infer from its own observations.

• 338 episodes, 24,911 frames (138 min at 3 Hz), 3 synchronised RGB cameras
• 39 episodes held out and marked in-band (is_test_holdout == 1)
• Real hardware, teleoperated demonstrations — no simulation

Authors

Collected, processed and released jointly by:

• Daye Lee (Addinedu) — @isYes · github.com/Daye-Lee18
• Gyujeong Song (Addinedu) — @songgyujeong

Task

Pick up the casting and place it into the bin on the AMR.

There is a single task string for the whole dataset. The scene factors below are deliberately not encoded in the instruction: a policy is meant to perceive them, not be told them. They are provided as separate columns so you can also train the conditioned variant and compare.

The factor grid

Factor	Values	Type
Item shape	circle, ellipse, square	categorical
Item pose	default, horizontal, vertical, tilted	categorical
Item side	center, left, right (where the item sits in the bin)	categorical
AMR position	PO_01 … PO_09, on a 3×3 grid (3.5 cm × 4 cm spacing)	continuous, not conditioned
AMR heading	−15°, 0°, +15°	continuous, not conditioned

The AMR grid is grouped into three distance bands by the measured robot-to-AMR distance: near (< 30 cm), middle (30–35 cm), far (> 35 cm).

Episode counts

Shape × pose:

	default	horizontal	vertical	tilted	total
circle	104	—	—	—	104
ellipse	—	63	—	36	99
square	—	63	36	36	135

Other margins: AMR distance near 153 / middle 152 / far 33; AMR heading 0° 134 / +15° 102 / −15° 102; item side center 230 / left 54 / right 54.

AMR positions PO_01–PO_06 were recorded at all three headings; the far row (PO_07–PO_09) was recorded at 0° only, 11 episodes each.

Data format

LeRobot dataset, codebase_version: v3.0 — parquet for the low-dimensional streams, AV1 mp4 for video.

Key	dtype	shape	Notes
observation.images.top	video	480×640×3	overhead view of the workspace
observation.images.side	video	480×640×3	side view of the arm
observation.images.wrist	video	480×640×3	wrist-mounted camera
observation.state	float32	(7,)	J1–J6 in degrees, gripper 0–100
action	float32	(7,)	the next follower joint position, same units
item_shape	int64	(1,)	0 circle · 1 ellipse · 2 square
item_pose	int64	(1,)	0 default · 1 horizontal · 2 vertical · 3 tilted
item_side	int64	(1,)	0 center · 1 left · 2 right
amr_position	int64	(1,)	1–9, the PO_NN grid cell
amr_distance	int64	(1,)	0 near · 1 middle · 2 far
amr_distance_cm	float32	(1,)	measured distance for that grid cell
amr_angle_deg	float32	(1,)	−15.0, 0.0 or 15.0
is_test_holdout	int64	(1,)	1 for the 39 held-out episodes

The factor columns are constant within an episode; they are stored per frame so that a single __getitem__ gives you the labels alongside the observation.

Note on action. Demonstrations were teleoperated with a leader–follower rig, but the leader signal was not retained: the action recorded here is the follower arm's own joint position one step later, i.e. action[t] == observation.state[t+1] exactly, within an episode. Training on this data is therefore next-state prediction rather than command regression, and the final frame of each episode has no successor (episode length is one less than the number of logged samples).

Cameras: top = Logitech C920e overhead, side = Logitech C920e, wrist = Sonix USB module on the gripper. Frames were captured at 30 Hz and stored at 3 Hz (every 10th frame) — at the action-chunk sizes used in our experiments, full temporal resolution was not needed and the subsample cuts storage roughly tenfold.

Episode length: 35–133 frames, mean 74, median 73 (≈12–44 s of wall-clock).

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("isYes/mycobot280_amr_pickplace")
print(ds.num_episodes, ds.num_frames, ds.fps)

sample = ds[0]
sample["observation.images.top"]   # (3, 480, 640) float32
sample["observation.state"]        # (7,)
sample["item_shape"], sample["item_pose"], sample["amr_angle_deg"]

Reading the labels without decoding video

Indexing ds[i] decodes three video frames, so never use it just to read a label. Go through ds.hf_dataset, which is the parquet table behind the dataset — the whole 338-episode label set comes back in a couple of seconds:

hf = ds.hf_dataset   # no video decoding

Reproducing the held-out split

The 39 held-out episodes cover every shape × pose combination and all three distance bands, and were excluded from both train and validation:

hold_eps = {int(e) for e, h in zip(hf["episode_index"], hf["is_test_holdout"]) if h == 1}
train_eps = [i for i in range(ds.num_episodes) if i not in hold_eps]
len(hold_eps)   # 39

Then load only what you want with the episodes= argument:

train_ds = LeRobotDataset("isYes/mycobot280_amr_pickplace", episodes=train_eps)

Building an OOD split along an unconditioned axis

AMR position and heading are continuous and were never used for conditioning, which makes them the natural OOD axes — e.g. train on near + middle and evaluate on far, or hold out one heading:

far_eps = {int(e) for e, d in zip(hf["episode_index"], hf["amr_distance"]) if d == 2}
len(far_eps)    # 33 episodes, 2739 frames

Limitations

Please read these before drawing conclusions from slices of this data.

1. Pose and position are confounded at far. The far row (PO_07–PO_09) was only recorded with default and horizontal items at 0° heading — no tilted or vertical episodes exist there. Pose is therefore not independent of position in the far band, and a far slice is not a clean pose-generalization test.
2. vertical is the thinnest cell (36 episodes, square only) and in our experiments it was consistently the hardest condition. Treat per-cell numbers there as directional.
3. far is small — 33 episodes total, 3 of which are in the held-out set.
4. One episode is missing: c_default_0deg_PO06_EP01 was dropped during collection, so the circle/default cell has 104 rather than the intended 105 episodes.
5. Demonstrations are teleoperated and not success-filtered by a formal criterion. They are operator-judged successful pick-and-place executions; there is no per-episode success score.
6. observations/qvel in the source HDF5 was all zeros and is not carried over.

Provenance

Collected on a myCobot 280 with a leader–follower teleoperation rig, logged as synchronised JPEG streams plus a follower-joint CSV, then assembled into per-episode HDF5 and converted to the LeRobot format. Collection, HDF5 assembly and the conversion script live in the project repository:

• Code: https://github.com/Daye-Lee18/VLA_exp
• Converter: scripts/data/hdf5_to_lerobot.py
• Collection: scripts/collect/1_collect_image_data.py, scripts/data/build_hdf5_v2.py

Original episode identifiers follow the pattern {shape}_{pose}_{deg}_PO{NN}_{arc}_EP{NN} and can be reconstructed from the label columns.

License

Released under the Apache 2.0 license. The upstream training code this dataset was built for is a fork of ACT / act-plus-plus (MIT, © 2023 Tony Z. Zhao).

Citation

@misc{mycobot280_amr_pickplace,
  title  = {mycobot280 AMR Pick-and-Place with Factored Shape/Pose/Viewpoint Variation},
  author = {Lee, Daye and Song, Gyujeong},
  institution = {Addinedu},
  year   = {2026},
  url    = {https://huggingface.co/datasets/isYes/mycobot280_amr_pickplace}
}