ITHwangg/svla_koch_pickplace_v2

Dataset Overview

• The dataset was created by the team Lebotica during LeRobot Worldwide Hackathon and used for training the SmolVLA model on structured robotic manipulation prompts
• The dataset consists of 90 tasks and 1 instruction, and there are the two types of episodes:
  • episode_0 ~ episode_53: Pick a color ball among the balls scattered on the white plate and place it in the corresponding color plate.
  • episode_54 ~ episode_89: Pick a color ball among the 9 balls placed at the fixed positions in the white plate and place it in the corresponding color plate.
• You can check the demo of the trained SmolVLA in the Hackathon Demo Page (Team number: 76).

Dataset Structure

├── data
│   └── chunk-000
│       ├── episode_000000.parquet
│       ├── ...
│       └── episode_000089.parquet
├── meta
│   ├── episodes.jsonl
│   ├── episodes_stats.jsonl
│   ├── info.json
│   └── tasks.jsonl
└── videos
    └── chunk-000
        ├── observation.images.side
        │   ├── episode_000000.mp4
        │   ├── ...
        │   └── episode_000089.mp4
        └── observation.images.top
            ├── episode_000000.mp4
            ├── ...
            └── episode_000089.mp4

• The tasks.json file contains an array of 90 task prompts. Each prompt follows a structured template for robotic manipulation.

• Example prompt:

  Pick a (red | blue | green) ball from the (top | middle | bottom)-(left | center | right) and place in the (red | blue | green) plate.
  

Usage

To use this dataset for training SmolVLA:

1. First, install the required dependencies:

   git clone https://github.com/huggingface/lerobot.git
   cd lerobot
   pip install -e ".[smolvla]"
   

2. Train SmolVLA

   python lerobot/scripts/train.py \
    --dataset.repo_id=ITHwangg/svla_koch_pickplace_v2 \
    --policy.path=lerobot/smolvla_base \
    --num_workers=8 \
    --batch_size=64 \
    --steps=100000 \
    --eval_freq=500 \
    --log_freq=10 \
    --save_freq=500 \
    --save_checkpoint=true
   

3. Caution

   • Currently, the python script refers to the branch named v2.1.
   • Every data/chunk-000/*.parquet has only the task index 0 so you should map epicode indexes to task indexes one by one:

     # lerobot/lerobot/common/datasets/lerobot_dataset.py
     class LeRobotDataset(torch.utils.data.Dataset):
        def __init__(
            self,
            repo_id: str,
            root: str | Path | None = None,
            episodes: list[int] | None = None,
            image_transforms: Callable | None = None,
            delta_timestamps: dict[list[float]] | None = None,
            tolerance_s: float = 1e-4,
            revision: str | None = None,
            force_cache_sync: bool = False,
            download_videos: bool = True,
            video_backend: str | None = None,
        ):
     
        ...
     
        # Load actual data
        try:
            if force_cache_sync:
                raise FileNotFoundError
            assert all((self.root / fpath).is_file() for fpath in self.get_episodes_file_paths())
            self.hf_dataset = self.load_hf_dataset()
        except (AssertionError, FileNotFoundError, NotADirectoryError):
            self.revision = get_safe_version(self.repo_id, self.revision)
            self.download_episodes(download_videos)
            self.hf_dataset = self.load_hf_dataset()
     
        # HERE ###########################
        # After loading the dataset and setting up episode_data_index
        if self.hf_dataset is not None:
            # Create a new column with task_index = episode_index
            new_task_index = torch.stack(self.hf_dataset["episode_index"])
            self.hf_dataset = self.hf_dataset.map(
                lambda x, idx: {"task_index": new_task_index[idx]}, with_indices=True
            )
        ##################################
     
        self.episode_data_index = get_episode_data_index(self.meta.episodes, self.episodes)
     
        ...
     

License

This dataset is released under the MIT License.