The dataset contains 10,123 episodes across 11 manipulation skills recorded from a first-person (ego-view) perspective in retail environments. Latent actions are encoded using LAPA (Latent Action Pretraining from Videos), a codebook-based action quantization model trained on large-scale egocentric video.

Dataset Summary

Dataset	Skill	Episodes	Frames
gt001.manipulation_cart_pushing	Cart Pushing	1,180	354,508
gt001.manipulation_grasping	Grasping	1,590	484,619
gt001.manipulation_holding	Holding	1,558	488,087
gt001.manipulation_holding_item	Holding Item	176	53,760
gt001.manipulation_lifting	Lifting	445	137,354
gt001.manipulation_object_manipulation_	Object Manipulation	188	55,303
gt001.manipulation_picking_up_item	Picking Up Item	1,550	473,327
gt001.manipulation_placing_item_in_basket	Placing Item in Basket	153	47,938
gt001.manipulation_placing_item_in_cart	Placing Item in Cart	423	137,514
gt001.manipulation_placing_item_on_shelf	Placing Item on Shelf	1,267	395,428
gt001.manipulation_reaching	Reaching	1,593	467,868
Total		10,123	3,095,706

Format

Codebase version: LeRobot v2.1
FPS: 30
Video: observation.images.ego_view — 640×480, H.264, egocentric (first-person)
Action: 4 latent action indices per frame, encoded by LAPA (codebook size 8, sequence length 4)
Annotations: Per-episode natural language task descriptions and task titles

Parquet Schema

Each episode parquet file contains the following columns:

Column	Type	Description
`timestamp`	float32	Time in seconds from episode start (`frame_index / fps`)
`frame_index`	int64	Frame index within the episode
`episode_index`	int64	Episode index within the dataset
`index`	int64	Global frame index across all episodes
`task_index`	int64	Index into `meta/tasks.jsonl`
`action`	int64[4]	LAPA latent action codes for this frame
`annotation.human.action.task_description`	string	Full natural language description of the task
`annotation.human.action.task_title`	string	Short task title

Dataset Structure

Each skill is a self-contained sub-directory following the LeRobot v2.1 layout:

gt001.manipulation_<skill>/
├── meta/
│   ├── info.json              # Dataset metadata (fps, features, total episodes)
│   ├── episodes.jsonl         # Per-episode metadata (length, task)
│   ├── episodes_stats.jsonl   # Per-episode statistics (mean/std/min/max per feature)
│   └── tasks.jsonl            # Task index → task description mapping
├── data/
│   └── chunk-000/
│       ├── episode_000000.parquet
│       ├── episode_000001.parquet
│       └── ...
└── videos/
    └── chunk-000/
        └── observation.images.ego_view/
            ├── episode_000000.mp4
            ├── episode_000001.mp4
            └── ...

Usage

With LeRobot

Install LeRobot:

pip install lerobot

Load a single skill dataset:

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

# Load one skill (downloads automatically)
dataset = LeRobotDataset(
    repo_id="DreamVu/Retail-VLA-10K",
    root="gt001.manipulation_grasping",  # sub-directory for the skill
)

print(f"Episodes: {dataset.num_episodes}")
print(f"Frames:   {len(dataset)}")

# Access a frame
frame = dataset[0]
print(frame.keys())
# dict_keys(['observation.images.ego_view', 'action', 'timestamp', ...])

Load multiple skills together:

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
from torch.utils.data import ConcatDataset

skills = [
    "gt001.manipulation_grasping",
    "gt001.manipulation_reaching",
    "gt001.manipulation_picking_up_item",
]

datasets = [
    LeRobotDataset("DreamVu/Retail-VLA-10K", root=skill)
    for skill in skills
]
combined = ConcatDataset(datasets)
print(f"Total frames: {len(combined)}")

With HuggingFace `datasets`

from datasets import load_dataset

# Load parquet-only (no videos) for a specific skill
ds = load_dataset("DreamVu/Retail-VLA-10K", name="grasping", split="train")
print(ds)
# Dataset with columns: timestamp, frame_index, episode_index, index,
#                       task_index, action, annotation.human.action.task_description,
#                       annotation.human.action.task_title

Download a single skill manually

huggingface-cli download DreamVu/Retail-VLA-10K \
    --repo-type dataset \
    --include "gt001.manipulation_grasping/**" \
    --local-dir ./retail_vla

License

This dataset is released under CC BY-NC 4.0. It is intended for non-commercial research use only.

Downloads last month: 4

You need to agree to share your contact information to access this dataset

Retail-VLA-10K