Cache-SCA/SO101-ma_distribute_chocolate_pie_each_plate_30fps

SO101 — Distribute Chocolate Pie to Each Plate (30 FPS)

LeRobot-format dataset of an SO-101 follower arm distributing two chocolate pies onto two plates, recorded at 30 FPS with top-view RealSense (RGB + depth) and a wrist-mounted UVC camera.

Task

place each chocolate pie on a plate

Two chocolate_pie objects must be picked and placed onto plate_1 and plate_2 respectively (one pie per plate). Initial workspace positions are randomized per batch via the --num-random-seeds mechanism.

Episode summary

Field	Value
Episodes	100
Frames	156,752 (~87 min trajectory @ 30 FPS)
FPS	30
Robot	SO-101 follower (5-DoF + gripper)
Cameras	top (Intel RealSense RGB-D, 640×480), wrist (Innomaker UVC, 640×480)
Codec	H.264
Median episode length	1,626 frames (~54 s)
Min / Max	863 / 1,989 frames

Features

Key	Type	Description
observation.state	(6,)	Motor positions: 5 arm joints + gripper
action	(6,)	Commanded joint positions
observation.images.realsense_topview	(480, 640, 3)	Top-view RGB
observation.images.wrist	(480, 640, 3)	Wrist-mounted RGB
observation.ee_pos.robot_xyzrpy	(6,)	End-effector pose in robot base frame
observation.gripper_binary	(1,)	Gripper open/close state
skill.type	str	Skill primitive (move_initial, move, move_and_open, gripper_close, …)
skill.progress	float	Within-skill normalized progress [0, 1]
skill.goal_position.joint	(6,)	Per-skill goal joint configuration
skill.goal_position.robot_xyzrpy	(6,)	Per-skill goal EE pose
skill.goal_position.gripper	(1,)	Per-skill goal gripper state
skill.verification_question	str	VLM-style verification prompt for this skill
subtask.natural_language	str	High-level subtask label (e.g. pick A and place at plate_1)
subtask.object_name	str	Target object name for the current subtask
subtask.target_position	(3,)	Target xyz for the subtask

How it was collected

The pipeline executes each episode as Forward + Judge + Reset in a closed loop:

1. Detection (Crop-then-Point) — a Gemini-3 Flash VLM identifies objects, draws bounding boxes, and proposes per-object grasp / interaction points (Turn 0/1/2). Pixels are lifted to robot frame via a Pix2Robot homography calibration.
2. Forward execution — a Manipulate-Anything (MA) closed-loop runner decomposes the instruction into sub-goals (object is in the gripper, object is placed on target, …) and executes them step-by-step using M2T2 grasp generation + Multi-IK selection. Each step is verified by a VLM sub-goal verifier.
3. Judge — a VLM compares the initial vs final top-view image to declare TRUE (success) / FALSE.
4. Reset — an LLM-generated Python script that picks each pie back to its initial scattered position. The episode is recorded into a separate reset dataset (not included here).

Across 20 random-seed batches × 5 episodes each, the scene is re-randomized so the model sees varied object placements.

Quick start

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("CoRL2026-CSI/SO101-ma_distribute_chocolate_pie_each_plate_30fps")
sample = dataset[0]
print(sample.keys())

For training:

python -m lerobot.scripts.train \
    --dataset.repo_id=CoRL2026-CSI/SO101-ma_distribute_chocolate_pie_each_plate_30fps \
    --policy.type=act \
    --output_dir=outputs/pie_distribute

Known limitations

• 13 internal episode slots originate from re-runs whose session-level artifacts were over-written before final dataset finalize; the trajectories themselves are valid demos (judge=TRUE), but their per-step Gemini logs in the on-disk session folder may not match those slots one-to-one.
• The wrist camera occasionally undergoes brief USB stream stalls during long sessions; per-episode dataset frames remain continuous, but the raw image timing may show small jitter at episode boundaries.

Hardware

• Arm: SO-101 follower
• Top camera: Intel RealSense D435 (serial 254622079503), 640×480 @ 30 FPS RGB + depth
• Wrist camera: Innomaker U20CAM, 640×480 @ 30 FPS, MJPG
• Calibration: Pix2Robot Charuco homography, mean reprojection error ≈ 2.6 mm

Citation

If you use this dataset, please cite:

@misc{so101_pie_distribute_2026,
  title  = {SO101 Distribute Chocolate Pies: A LeRobot Closed-Loop Demonstration Dataset},
  author = {Cho, Hyeonseok and contributors},
  year   = {2026},
  url    = {https://huggingface.co/CoRL2026-CSI/SO101-ma_distribute_chocolate_pie_each_plate_30fps}
}