CoRL2026-CSI/SO101-cap_stack_RGBblock_on_bluedish_10fps

SO101 CAP Stack RGB Blocks on Blue Dish

This dataset contains 100 LeRobot v3.0 demonstration episodes for an SO101 follower robot. The task is: Stack red, green, and blue blocks on the blue dish from bottom to top. The dataset was collected at 10 Hz and includes paired top-view and wrist-view RGB videos, robot state/action trajectories, and CAP skill annotations.

Dataset Details

Field	Value
Repository	CoRL2026-CSI/SO101-cap_stack_RGBblock_on_bluedish_10fps
LeRobot codebase version	v3.0
Robot type	so101_follower
FPS	10
Episodes	100
Frames	87074
Tasks	1
Split	0:100 train
Objects	red, green, and blue blocks; blue dish
Parquet files	9
Video files	20

Task And Annotations

Task: Stack red, green, and blue blocks on the blue dish from bottom to top.

Task indices: 0

Observed subtask annotations:

• pick red block and place on blue dish
• pick green block and place on red block
• pick blue block and place on green block

Representative skill types:

• move_initial
• move_and_open
• move
• gripper_close
• gripper_open
• move_and_close
• move_free

Representative skill-language annotations:

• move to initial state
• Approach red block and open gripper
• Pick up the red block
• grasp red block
• Lift the red block
• Move red block above blue dish
• Place red block on blue dish
• release object on blue dish
• Retreat from blue dish and close gripper
• Approach green block and open gripper
• Pick up the green block
• grasp green block

Observation And Action Space

Feature	Shape	Names / Notes
observation.state	6	shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos
action	6	shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos
observation.images.top	480x640x3	RGB video, 10 fps
observation.images.left_wrist	480x640x3	RGB video, 10 fps

Raw camera keys are observation.images.top and observation.images.left_wrist. The local SmolVLA training scripts map them to observation.images.camera2 and observation.images.camera1, respectively.

Files

meta/info.json
meta/tasks.parquet
meta/episodes/chunk-*/file-*.parquet
data/chunk-*/file-*.parquet
videos/{observation.images.top,observation.images.left_wrist}/chunk-*/file-*.mp4

The dataset uses the LeRobot v3.0 format. Episode metadata and frame-level trajectories are stored in parquet files; image observations are stored as H.264 MP4 videos referenced by the frame records.

Annotation Columns

Column	Shape
skill.natural_language	1
skill.verification_question	1
skill.type	1
skill.progress	1
skill.goal_position.joint	6
skill.goal_position.robot_xyzrpy	6
skill.goal_position.gripper	1
subtask.natural_language	1
subtask.object_name	1
subtask.target_position	3

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("CoRL2026-CSI/SO101-cap_stack_RGBblock_on_bluedish_10fps")
sample = dataset[0]

Intended Uses

This dataset is intended for robot imitation learning, action-chunking policy training, skill-conditioned behavior analysis, and reproducible LeRobot/SmolVLA experiments on the specified tabletop task.

Limitations

The dataset is task-specific and collected in a fixed workspace. It does not include an official validation or test split, nor does it include benchmark success-rate results. Downstream users should verify camera calibration, action normalization, and task-language assumptions before transferring policies to a different robot, workspace, or object set.