Dimios45/openpi-fold-towel-rollout-ablation

OpenPI MolmoAct2 fold-towel rollout ablation

Merged LeRobot v3 dataset: Dimios45/openpi-fold-towel-rollout-ablation.

This dataset contains the hardware inference rollouts collected with openpi-control rollout on the bimanual YAM cell. Every source run contributes all saved episodes, including partial episodes saved after Ctrl-C. The y/n labels are the operator's task-success labels.

Contents

• 21 episodes and 70,712 frames
• Nominal recording rate: 30 FPS
• Three video views per frame: top, left wrist, right wrist
• State/action vector: 14 values (6 joints + gripper per arm)
• Canonical LeRobot metadata under meta/ and trajectory data under data/
• Per-episode configuration/provenance: meta/rollout_metadata.jsonl
• Original recorder manifests: provenance/

Configurations tested

The effective duration is computed from the implementation's interpolation rule: ceil((chunk_size - 1) / speed) + 1 control ticks.

Source run	Speed	Chunk size	Approx. chunk duration	Episodes	Success labels
fold-towel	0.5	30	1.97 s	3	nnn
fold-towel-speed05-chunk15	0.5	15	0.97 s	3	nnn
fold-towel-speed05-chunk20	0.5	20	1.30 s	3	nnn
fold-towel-speed07-chunk30	0.7	30	1.43 s	3	nnn
fold-towel-speed08-chunk15	0.8	15	0.63 s	3	nnn
fold-towel-speed08-chunk30	0.8	30	1.27 s	3	nnn
fold-towel-speed10-chunk15	1	30	1.00 s	3	nnn

Episode mapping

The original recorder stores the natural-language prompt, but not the towel color. The object names below follow the operator's episode-order annotation for these runs:

Local episode slot	Object annotation	Prompt in dataset
1	pink towel	fold the towel
2	blue towel	fold the towel
3	black T-shirt	fold the tshirt

Important notes

• The original source manifests report all 21 episodes as unsuccessful (n).
• The object annotation is provenance supplied by the operator; it is not part of the original per-frame task string.
• The directory fold-towel-speed10-chunk15 is named as chunk 15, but its source manifest records speed=1.0 and chunk_size=30; the manifest is treated as the ground truth.
• Dataset timestamps use the declared 30 FPS. The saved episodes are about 110–115 nominal seconds each even when the wall-clock limit was 120 seconds.
• The rollout runtime used the PyAV video backend. This dataset preserves the recorded videos and does not recompress them.

Loading

The dataset is structured as a standard LeRobot v3 dataset. The sidecar configuration can be joined by global_episode_index:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("Dimios45/openpi-fold-towel-rollout-ablation")

The trajectory rows retain the original prompt in LeRobot's task field. Use meta/rollout_metadata.jsonl when filtering by speed, chunk size, object annotation, success label, or source run.