Dexora Real World Dataset

license: mit

Dexora: Open-Source VLA for High-DoF Bimanual Dexterity

🔥 News & Updates

• 2025-12-03: Released the full Real-World Dataset (12.2K episodes) on Hugging Face.
• 2025-12-12: Released the task-level view of the Real-World Dataset (one folder per high-level task) on Hugging Face.
• Coming soon: We will open-source the full 100K-episode simulation dataset.

--

📊 Dataset Overview

The Dexora corpus combines high-fidelity real-world teleoperation data with a large-scale simulated dataset designed to match the embodiment of the physical robot.

A. Dexora Real-World Dataset (High-Fidelity)

The Dexora real-world dataset consists of 12.2K teleoperated episodes, 2.92M frames, and 40.5 hours of data. Demonstrations are collected using a hybrid teleoperation system that couples an Exoskeleton (for arm control) with Vision Pro (for dexterous hand control), enabling precise 36-DoF bimanual manipulation on real hardware.

Video 1. Synchronized Multi-View Recordings. High-resolution streams from four synchronized views — an ego-centric head-mounted camera, left and right wrist-mounted cameras, and a static third-person scene camera — synchronized with 36-DoF robot proprioception.

Fig 1. High-Fidelity Real-World Scenes. Collected via our hybrid teleoperation system (Exoskeleton for arm + Vision Pro for hand), this dataset covers 347 objects across diverse environments. It captures varying lighting conditions, background clutter, and precise bimanual interactions essential for robust policy learning. Panels (a–d) correspond to four task categories: pick-and-place, assembly, articulation, and dexterous manipulation.

Fig 2. Task Categories & Action Distribution. Unlike standard gripper datasets, Dexora emphasizes high-DoF dexterity. The real-world data distribution includes Dexterous Manipulation (20%) (e.g., Twist Cap, Use Pen, Cut Leek) and Assembly (15%) (e.g., Separate Nested Bowls, Stack Ring Blocks), in addition to Articulated Objects (10%) and Pick-and-Place (55%).

Both the episodes and annotations follow the LIBERO-2.1 standard, including synchronized RGB observations, robot proprioception, actions, and language instructions.

Object Inventory & Reproducibility

Reproducibility is a core value of Dexora. To enable other labs and industry teams to faithfully recreate our environments, we release a curated object inventory that mirrors the physical setup used in our real-world experiments.

• Scale: 347 objects across 17 semantic categories (e.g., tools, containers, articulated objects, deformables, daily-use items).
• Coverage: Objects are chosen to stress dexterous control, bimanual coordination, and long-horizon manipulation.
• Procurement: Every item is linked to Taobao and/or Amazon, so researchers can rebuild the setup with minimal effort.

📑 Access Dexora Real-world Item List (Google Sheet)

Inventory Metadata Schema

The released Google Sheet follows the schema below:

You can filter by task type, category, or store to design controlled benchmarks or new task suites on top of Dexora.

B. Dexora Simulation Dataset (Large-Scale)

The Dexora simulation dataset contains 100K episodes generated in MuJoCo, using the same 36-DoF dual-arm, dual-hand embodiment as the real robot. It provides large-scale, embodiment-matched experience focused on core skills such as pick-and-place, assembly, and articulation, which can be used for pre-training basic competence before fine-tuning on the real-world dataset.

Summary Statistics (Sim vs Real)

📂 Data Structure

Dexora follows the LIBERO-2.1 dataset standard. Each episode is stored as a self-contained trajectory with:

• Observations: multi-view RGB (and optionally depth), segmentation masks (when available).
• Robot State: joint positions/velocities for dual arms and dual hands, gripper/hand states.
• Actions: low-level control commands compatible with 36-DoF bimanual control.
• Language: High-level task descriptions. We provide 5 diverse natural language instructions per task, distributed evenly across all trajectories to enhance linguistic diversity.

We provide an additional task-level view (one folder per high-level task) on Hugging Face, alongside the original episode-centric LIBERO-2.1 layout. The latest complete structure is:

Dexora_Real-World_Dataset
├── airbot_articulation
│   ├── data
│   │   ├── chunk-000
│   │   │   ├── episode_000000.parquet
│   │   │   ├── episode_000001.parquet
│   │   │   ├── ...
│   │   ├── chunk-001
│   │   ├── ...
│   ├── videos
│   │   ├── chunk-000
│   │   │   ├── observation.images.front
│   │   │   │   ├── episode_000000.mp4
│   │   │   │   ├── episode_000001.mp4
│   │   │   │   ├── ...
│   │   ├── chunk-001
│   │   ├── ...
│   ├── meta
│   │   ├── info.json
│   │   ├── episodes.jsonl
│   │   ├── episodes_stats.jsonl
│   │   ├── modality.json
│   │   ├── stats.json
│   │   ├── tasks.jsonl
├── airbot_assemble
│   └── ...
├── airbot_dexterous
│   └── ...
├── airbot_pick_and_place
│   └── ...
├── task_level_episodes
│   ├── apply_tape_to_bottle
│   ├── arrange_apple_peach_pear
│   ├── fold_towel_bimanual
│   ├── move_toy_cars_from_plate_to_table
│   ├── ...
└── README.md

> **Note**: The exact folder names and file formats may be updated as we finalize the public release, but the overall **episode-centric LIBERO-2.1 structure** will be preserved.


[meta/info.json](meta/info.json):
    

      

        json
        
          Copy
        
      
      
{
    "codebase_version": "v2.1",
    "robot_type": "airbot_play",
    "total_episodes": 11517,
    "total_frames": 2919110,
    "total_tasks": 201,
    "total_videos": 46068,
    "total_chunks": 12,
    "chunks_size": 1000,
    "fps": 20,
    "splits": {
        "train": "0:2261"
    },
    "data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
    "video_path": "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4",
    "features": {
        "observation.images.top": {
            "dtype": "video",
            "shape": [
                480,
                640,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "video.height": 480,
                "video.width": 640,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 20,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.images.wrist_left": {
            "dtype": "video",
            "shape": [
                480,
                640,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "video.height": 480,
                "video.width": 640,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 20,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.images.wrist_right": {
            "dtype": "video",
            "shape": [
                480,
                640,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "video.height": 480,
                "video.width": 640,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 20,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.images.front": {
            "dtype": "video",
            "shape": [
                480,
                640,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "video.height": 480,
                "video.width": 640,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 20,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.state": {
            "dtype": "float32",
            "shape": [
                39
            ],
            "names": [
                "left_arm_joint_1",
                "left_arm_joint_2",
                "left_arm_joint_3",
                "left_arm_joint_4",
                "left_arm_joint_5",
                "left_arm_joint_6",
                "right_arm_joint_1",
                "right_arm_joint_2",
                "right_arm_joint_3",
                "right_arm_joint_4",
                "right_arm_joint_5",
                "right_arm_joint_6",
                "left_hand_joint_1",
                "left_hand_joint_2",
                "left_hand_joint_3",
                "left_hand_joint_4",
                "left_hand_joint_5",
                "left_hand_joint_6",
                "left_hand_joint_7",
                "left_hand_joint_8",
                "left_hand_joint_9",
                "left_hand_joint_10",
                "left_hand_joint_11",
                "left_hand_joint_12",
                "right_hand_joint_1",
                "right_hand_joint_2",
                "right_hand_joint_3",
                "right_hand_joint_4",
                "right_hand_joint_5",
                "right_hand_joint_6",
                "right_hand_joint_7",
                "right_hand_joint_8",
                "right_hand_joint_9",
                "right_hand_joint_10",
                "right_hand_joint_11",
                "right_hand_joint_12",
                "head_joint_1",
                "head_joint_2",
                "spine_joint"
            ]
        },
        "action": {
            "dtype": "float32",
            "shape": [
                39
            ],
            "names": [
                "left_arm_joint_1",
                "left_arm_joint_2",
                "left_arm_joint_3",
                "left_arm_joint_4",
                "left_arm_joint_5",
                "left_arm_joint_6",
                "right_arm_joint_1",
                "right_arm_joint_2",
                "right_arm_joint_3",
                "right_arm_joint_4",
                "right_arm_joint_5",
                "right_arm_joint_6",
                "left_hand_joint_1",
                "left_hand_joint_2",
                "left_hand_joint_3",
                "left_hand_joint_4",
                "left_hand_joint_5",
                "left_hand_joint_6",
                "left_hand_joint_7",
                "left_hand_joint_8",
                "left_hand_joint_9",
                "left_hand_joint_10",
                "left_hand_joint_11",
                "left_hand_joint_12",
                "right_hand_joint_1",
                "right_hand_joint_2",
                "right_hand_joint_3",
                "right_hand_joint_4",
                "right_hand_joint_5",
                "right_hand_joint_6",
                "right_hand_joint_7",
                "right_hand_joint_8",
                "right_hand_joint_9",
                "right_hand_joint_10",
                "right_hand_joint_11",
                "right_hand_joint_12",
                "head_joint_1",
                "head_joint_2",
                "spine_joint"
            ]
        },
        "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
        }
    }
}


  
🤗
Data Source: Hugging Face ↗
🔄 Daily sync (03:00 UTC)
AI Summary: Based on Hugging Face metadata. Not a recommendation.

📊 FNI Methodology
📚 Knowledge Baseℹ️ Verify with original source
 
 
 
 
 🛡️ Dataset Transparency Report 
 
Verified data manifest for traceability and transparency.
 
 
 
 
100% Data Disclosure Active
 
 
 
 
 
 🆔 Identity & Source 
 
 
 
 id 
 
hf-dataset--dexora--dexora_real-world_dataset
 
 
 source 
 
huggingface
 
 
 author 
 
Dexora
 
 
 tags 
 
license:mitmodality:videoregion:us
 
 
 
 
 ⚙️ Technical Specs 
 
 
 
 architecture 
 
null
 
 
 params billions 
 
null
 
 
 context length 
 
null
 
 
 
 
 📊 Engagement & Metrics 
 
 
 
 likes 
 
20
 
 
 downloads 
 
38,080
 
 
 
 
 
 
Free2AITools Constitutional Data Pipeline: Curated disclosure mode active. (V15.x Standard)