SMART-Data

SMART-Data is a large-scale synthetic dataset for articulated-object manipulation, generated entirely in simulation with a scalable synthesis pipeline.
It covers five robot setups and a hierarchical taxonomy of atomic, composite, and long-horizon tasks, spanning 23 manipulation skills, 44 task types, and diverse articulated objects with revolute, prismatic, and compound joints.
The released demonstrations are organized in LeRobot v3.0 format for robot learning and VLA pretraining.

🔭 Dataset Overview

Statistics of SMART-Data, including robot-setup distribution, task complexity, and articulated-object coverage.

✨ Dataset Features

• Large-scale synthetic data: SMART-Data contains over 1M simulated manipulation episodes, approximately 500M frames, and over 4,600 hours of demonstrations.
• Five robot setups: The dataset covers ARX AC One, R1Pro, Dual RM75, Dual Franka, and Single Franka, providing demonstrations from both single-arm and dual-arm embodiments.
• Hierarchical task design: Tasks are organized into atomic, composite, and long-horizon categories, supporting primitive skills as well as multi-stage compositional reasoning.
• Broad skill coverage: The dataset spans 23 manipulation skills and 44 task types across approximately 10K task YAML configurations.
• Diverse articulated objects: It covers 23 of the 88 functional object categories in the SMART-Sim asset library, including 2,507 articulated objects with revolute, prismatic, and compound joints.
• Rich interaction diversity: Demonstrations vary in object pose, robot initialization, camera extrinsics, visual appearance, and physical properties.
• Multi-view observations: The data include global and arm-specific camera views for learning manipulation policies from complementary perspectives.
• Scalable scene coverage: The released trajectories are collected across 1,122 distinct scenes.
• VLA-ready format: All demonstrations are stored in LeRobot v3.0 format for behavior cloning, VLA pretraining, and downstream manipulation research.

🏁 Get Started

⬇️ Download

# Make sure you have Git LFS installed
git lfs install

# For private repositories, use a Hugging Face access token with read access
git clone https://huggingface.co/datasets/TeleEmbodied/SMART-Data

# Clone without downloading the large files, leaving only their pointers
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/TeleEmbodied/SMART-Data

🗂️ Dataset Structure

The following example shows the fully expanded dataset layout:

SMART-Data/
├── AC1/                                             # robot platform
│   ├── atomic/                                      # primitive tasks
│   │   ├── left/
│   │   │   └── left_close_dishwasher/               # task
│   │   │       ├── data/
│   │   │       │   └── chunk-000/
│   │   │       │       ├── file-000.parquet
│   │   │       │       ├── file-001.parquet
│   │   │       │       ├── file-002.parquet
│   │   │       │       └── ...
│   │   │       ├── meta/
│   │   │       │   ├── episodes/
│   │   │       │   │   └── chunk-000/
│   │   │       │   │       ├── file-000.parquet
│   │   │       │   │       ├── file-001.parquet
│   │   │       │   │       ├── file-002.parquet
│   │   │       │   │       └── ...
│   │   │       │   ├── info.json
│   │   │       │   ├── norm_stats.json
│   │   │       │   ├── stats.json
│   │   │       │   ├── tasks.parquet
│   │   │       │   └── ...
│   │   │       ├── videos/
│   │   │       │   ├── observation.images.camera_global/
│   │   │       │   │   └── chunk-000/
│   │   │       │   │       ├── file-000.mp4
│   │   │       │   │       ├── file-001.mp4
│   │   │       │   │       └── ...
│   │   │       │   ├── observation.images.camera_left/
│   │   │       │   │   └── ...
│   │   │       │   ├── observation.images.camera_right/
│   │   │       │   │   └── ...
│   │   │       │   └── ...
│   │   │       └── ...
│   │   ├── right/
│   │   │   └── ...
│   │   └── ...
│   ├── composite/                                   # multi-skill tasks
│   │   └── ...
│   └── long_horizon/                                # long-horizon tasks
│       └── ...
├── Franka_dual/
│   └── ...
├── Franka_single/
│   └── ...
├── R1Pro/
│   └── ...
└── RM75/
    └── ...

📄 Data Format

<task>/                         # LeRobot dataset
├── data/                       # frame-level data
│   └── chunk-000/
│       └── file-000.parquet    # action, state, and timestamps
├── meta/                       # dataset metadata
│   ├── info.json               # schema and dataset information
│   ├── norm_stats.json
│   ├── stats.json
│   ├── tasks.parquet           # task instruction
│   └── episodes/               # episode metadata
└── videos/                     # camera streams
    ├── observation.images.camera_global/
    ├── observation.images.camera_left/
    └── observation.images.camera_right/

🔎 Example meta/info.json

{
  "codebase_version": "v3.0",
  "robot_type": "AC1",
  "total_episodes": 500,
  "total_frames": 170756,
  "total_tasks": 1,
  "chunks_size": 1000,
  "data_files_size_in_mb": 100,
  "video_files_size_in_mb": 200,
  "fps": 30,
  "splits": {
    "train": "0:500"
  },
  "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
  "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
  "features": {
    "observation.images.camera_global": {
      "dtype": "video",
      "shape": [540, 960, 3],
      "names": ["height", "width", "rgb"],
      "info": {
        "video.height": 540,
        "video.width": 960,
        "video.codec": "h264",
        "video.pix_fmt": "yuv420p",
        "video.is_depth_map": false,
        "video.fps": 30,
        "video.channels": 3,
        "has_audio": false
      }
    },
    "observation.images.camera_left": {
      "dtype": "video",
      "shape": [360, 640, 3],
      "names": ["height", "width", "rgb"],
      "info": {
        "video.height": 360,
        "video.width": 640,
        "video.codec": "h264",
        "video.pix_fmt": "yuv420p",
        "video.is_depth_map": false,
        "video.fps": 30,
        "video.channels": 3,
        "has_audio": false
      }
    },
    "observation.images.camera_right": {
      "dtype": "video",
      "shape": [360, 640, 3],
      "names": ["height", "width", "rgb"],
      "info": {
        "video.height": 360,
        "video.width": 640,
        "video.codec": "h264",
        "video.pix_fmt": "yuv420p",
        "video.is_depth_map": false,
        "video.fps": 30,
        "video.channels": 3,
        "has_audio": false
      }
    },
    "action": {
      "dtype": "float32",
      "shape": [14],
      "names": {
        "motors": [
          "left_arm_0", "left_arm_1", "left_arm_2",
          "left_arm_3", "left_arm_4", "left_arm_5",
          "right_arm_0", "right_arm_1", "right_arm_2",
          "right_arm_3", "right_arm_4", "right_arm_5",
          "left_gripper", "right_gripper"
        ]
      }
    },
    "observation.state": {
      "dtype": "float32",
      "shape": [14],
      "names": [
        "left_arm_0", "left_arm_1", "left_arm_2",
        "left_arm_3", "left_arm_4", "left_arm_5",
        "right_arm_0", "right_arm_1", "right_arm_2",
        "right_arm_3", "right_arm_4", "right_arm_5",
        "left_gripper", "right_gripper"
      ]
    },
    "timestamp": {
      "dtype": "float32",
      "shape": [1],
      "names": null
    },
    "frame_index": {
      "dtype": "int64",
      "shape": [1],
      "names": null
    },
    "episode_index": {
      "dtype": "int64",
      "shape": [1],
      "names": null
    },
    "index": {
      "dtype": "int64",
      "shape": [1],
      "names": null
    },
    "task_index": {
      "dtype": "int64",
      "shape": [1],
      "names": null
    }
  }
}