LeRobot-SO101-Pick-Place

Dataset Summary

LeRobot dataset for SO101 pick-and-place with three sponge manipulation tasks and fixed/random object layouts. The dataset is intended for imitation learning, behavior cloning, policy evaluation, and robustness studies under layout changes.

Supported Tasks

• Imitation learning
• Behavior cloning
• Policy evaluation
• Robustness to layout changes

Dataset Structure

The dataset is organized into three task groups:

• task1-single-sponge-no-distractors
• task2-multiple-sponges-no-distractors
• task3-single-sponge-with-distractors

Each task contains two location splits:

• fixed-locations: layouts drawn from 6 fixed workspace locations
• random-locations: layouts with randomized locations

Each split keeps the LeRobot directory structure:

• data/
• meta/
• videos/

Visualizable Child Datasets

Each split is also published as its own root-level LeRobot dataset repo so it can be used directly with the LeRobot visualizer:

• LeRobot-SO101-task1-single-sponge-no-distractors-fixed-locations
• LeRobot-SO101-task1-single-sponge-no-distractors-random-locations
• LeRobot-SO101-task2-multiple-sponges-no-distractors-fixed-locations
• LeRobot-SO101-task2-multiple-sponges-no-distractors-random-locations
• LeRobot-SO101-task3-single-sponge-with-distractors-fixed-locations
• LeRobot-SO101-task3-single-sponge-with-distractors-random-locations

Task Definitions

task1-single-sponge-no-distractors

Single sponge pick-and-place with no distractor objects.

task2-multiple-sponges-no-distractors

Multiple sponge pick-and-place with no distractor objects.

task3-single-sponge-with-distractors

Single sponge pick-and-place with distractor objects present in the scene.

Split Details

fixed-locations

This split corresponds to the previously named known split. It uses fixed workspace layouts from 6 predefined locations.

random-locations

This split corresponds to the previously named unknown split. It uses randomized workspace layouts.

Episode Distribution

Task 1

• fixed-locations: 64 night episodes
• random-locations: 49 night episodes, 15 morning episodes

Task 2

• fixed-locations: 32 morning episodes, 32 night episodes
• random-locations: 32 morning episodes, 32 night episodes

Task 3

• fixed-locations: 32 morning episodes, 32 night episodes
• random-locations: 32 morning episodes, 32 night episodes

Data Collection

The dataset was collected on a real SO101 leader robot. Observations include two camera views:

• overhead camera
• wrist camera

Task prompts are in English. The dataset card currently leaves the exact recorded action/control representation unspecified.

Intended Use

This dataset is intended for:

• imitation learning
• behavior cloning
• policy evaluation
• robustness experiments under location changes

Limitations

• only 3 tasks are included
• only sponge objects are used
• camera placement is fully fixed
• lighting conditions are limited to the collected morning and night settings

Creator

Created by Aswinkumar.

• Website: aswinkumar.me

Citation

If you use this dataset, please cite the dataset repository and credit the creator:

@misc{aswinkumar_lerobot_so101_pick_place,
  title        = {LeRobot-SO101-Pick-Place},
  author       = {Aswinkumar},
  howpublished = {\url{https://huggingface.co/datasets/aswinkumar99/LeRobot-SO101-Pick-Place}},
  year         = {2026}
}