NandithaaaG/grab_place_training

SO101 Grab & Place Dataset

Overview

This dataset contains robot demonstrations collected using the SO101 robotic arm for a Grab & Place manipulation task.

The demonstrations were collected through robot teleoperation using the LeRobot framework.

The dataset is intended for robot imitation learning, physical AI experimentation, and robotic manipulation research.

Dataset Statistics

Property	Details
Robot	SO101
Task	Grab & Place
Total Episodes	20
Total Frames	29,917
Recording FPS	30 FPS
Camera Views	Front + Side
Image Resolution	640 x 480
Robot State Dimension	6
Action Dimension	6
Robot Type	so_follower
Dataset Format	LeRobot

Task Description

The task consists of manipulating an object using the SO101 robotic arm.

A typical demonstration includes:

1. Observing the target object.
2. Moving the robotic arm toward the object.
3. Grasping the object using the gripper.
4. Moving the object toward the target box.
5. Placing the object inside the box.

Camera Observations

The dataset contains two camera streams.

Front Camera

observation.images.front

Resolution: 640 x 480 x 3

Frame rate: 30 FPS

Side Camera

observation.images.side

Resolution: 640 x 480 x 3

Frame rate: 30 FPS

Robot State

Each observation contains six robot joint positions:

• shoulder_pan.pos
• shoulder_lift.pos
• elbow_flex.pos
• wrist_flex.pos
• wrist_roll.pos
• gripper.pos

Action

The action space contains six joint position values:

• shoulder_pan.pos
• shoulder_lift.pos
• elbow_flex.pos
• wrist_flex.pos
• wrist_roll.pos
• gripper.pos

Dataset Contents

The dataset contains:

• Robot joint states
• Robot actions
• Front camera videos
• Side camera videos
• Timestamps
• Frame indices
• Episode indices
• Task indices
• Episode metadata
• Dataset statistics

Imitation Learning

This dataset can be used to train a robot policy from demonstrations.

The learning pipeline is:

Front Camera + Side Camera + Robot State -> Policy -> Robot Action

The dataset was prepared for experimentation with Action Chunking with Transformers (ACT) using LeRobot.

Training Configuration

• Policy: ACT
• Vision Backbone: ResNet18
• Robot: SO101
• Camera Views: Front + Side
• State Dimension: 6
• Action Dimension: 6

Data Collection

A total of 20 demonstration episodes were collected.

The dataset contains:

• 20 Episodes
• 29,917 Frames
• 30 FPS
• 2 Camera Views
• 6 Robot State Values
• 6 Action Values

Intended Applications

This dataset can be used for:

• Robot imitation learning
• Robotic manipulation
• Physical AI research
• Vision-based robot control
• ACT policy training
• SO101 experimentation
• Robot learning from demonstrations

Limitations

The dataset contains 20 demonstration episodes from a specific Grab & Place setup.

Performance may vary with changes in object position, object appearance, lighting, camera viewpoint, robot initial position, object placement position, and environment configuration.

Additional demonstrations with greater task variation may improve generalization.

Hardware

Robot: SO101

Robot Type: so_follower

Software

• LeRobot
• Hugging Face
• Python
• ACT

Dataset Metadata

• Total Episodes: 20
• Total Frames: 29,917
• Total Tasks: 1
• FPS: 30
• Robot Type: so_follower

Citation

Nanditha G. SO101 Grab & Place Dataset Hugging Face Dataset Repository 2026

Acknowledgements

This dataset was collected using the SO101 robotic arm and prepared using the LeRobot framework.