cueng/so101_demo_bowl

SO101 Put [object] into Bowl

Dataset Summary

This dataset is formulated for use with the LeRobot framework. It contains robotic manipulation demonstrations for the SO101 embodiment performing a standardized pick-and-place task: "Put [object] into a bowl." The dataset captures precise manipulation trajectories across 10 distinct household and office objects against a fixed background, recorded from multiple camera angles. It is designed to train and evaluate visuomotor policies for robotic manipulation. Recorded and Validated by Sorrawit Poomseetong

---

Dataset Details

Statistics

• Robot: SO101 Follower
• Total Rows/Frames: 221,896
• Control Frequency: 30 Hz
• Demonstrations: 100 per object (1,000 total episodes)
• Background Setup: Fixed background
• Action Horizon/Length: ~200 steps per episode

Sensor Configuration

The dataset includes synchronized multi-view RGB image observations:

• Top Camera: (observation.images.top) Provides a fixed, global overview of the workspace, the bowl, and the object's initial position.
• Wrist Camera: (observation.images.wrist) Provides an ego-centric view from the robot's end-effector, crucial for fine manipulation and grasping tasks.

---

Task & Objects

Task Instruction: "Put [object] into a bowl."

The dataset covers variations of this task across the following 10 objects. Each object features exactly 100 successful demonstrations:

• A plastic spoon
• A metal spoon
• A scissors
• A roll of tissue paper
• A blue marker
• A black marker
• A nipper
• A glue
• A white tape
• A screw driver

---

Data Structure

Data is formatted natively for lerobot and contains the following standard feature keys per step:

Observations

• observation.images.top (image): RGB image from the overhead camera.
• observation.images.wrist (image): RGB image from the wrist camera.
• observation.state (list of float32): The state of the SO101 robot.

Actions & Metadata

• action (List[float32]): The commanded actions at 30 Hz.
• timestamp (float32): Time elapsed in seconds since the start of the episode.
• frame_index (int64): Frame number within the current episode.
• episode_index (int64): Index of the demonstration episode (0 to 999).
• index (int64): Global index of the frame.
• task_index (int64): Corresponding ID for the specific object task.

---

Usage with LeRobot

You can load this dataset directly using the LeRobot library:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

# Load the dataset
dataset = LeRobotDataset("cueng/so101_demo_bowl")

# Load first frame
frame = dataset[0]
print(frame.keys())