README.md

---
license: mit
dataset_info:
  features:
    - name: action
      dtype: numpy
    - name: observation.state
      dtype: numpy
    - name: observation.images.phone
      dtype: video
    - name: observation.images.gripper
      dtype: video
  splits:
    - name: train
      num_examples: 100
  task_categories:
    - robotics
tags:
  - robotics
  - augmented
  - robot-manipulation
  - data-augmentation
---

# LeRobot Augmented Dataset

This dataset is an augmented version of the original LeRobot dataset. The augmentation expands the dataset by creating 4 versions of each original episode:

1. **Original data** - preserved as-is
2. **Horizontally flipped images** - original action/state vectors
3. **Shoulder pan negated** - original images with shoulder pan values negated in action/state vectors
4. **Both flipped and negated** - horizontally flipped images with negated shoulder pan values

## Augmentation Process

The augmentation process quadruples the size of the dataset by creating three variants of each original episode:

- **Image flipping**: Horizontally flipping camera frames to simulate the robot operating from the opposite side
- **Action negation**: Negating shoulder pan values to simulate opposite directional movement
- **Combined augmentation**: Both flipping images and negating shoulder pan values

This 4x expansion provides more diverse training data for robotic control tasks, helping models generalize better to different perspectives and movement directions.

## Original Task

The dataset contains episodes where a robot is performing the task: "Pick up the black pen and place it in the mug."

## Structure

The dataset maintains the same structure as the original:
- `data/`: Contains episode data in parquet format (4x the original amount)
- `videos/`: Contains video recordings of episodes (4x the original amount)
- `meta/`: Contains metadata about the dataset