license: mit
Dexora: Open-Source VLA for High-DoF Bimanual Dexterity
🔥 News & Updates
- 2025-12-03: Released the full Real-World Dataset (10K episodes) on Hugging Face.
--
📊 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 11.5K 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 ego-centric, third-person, and wrist-mounted cameras, 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:
| Column | Description |
|---|---|
| Object Name (EN & CN) | Bilingual identification for global researchers. |
| Task Type | One of: pick-and-place, assemble, articulation, dexterous. |
| Purchase Link | Direct links to Taobao & Amazon for easy procurement and restock. |
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)
| Split | Episodes | Frames | Hours (approx.) | Task Types |
|---|---|---|---|---|
| Simulated | 100K | 6.5M | TBD | Pick-and-place, assembly, articulation |
| Real-World | 10K | 3.2M | 177.5 | Teleoperated bimanual tasks with high-DoF hands, cluttered scenes, fine-grain object interactions |
📂 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.
An example high-level directory layout is:
data
├── real
│ ├── articulation
│ │ └── ...
│ ├── assembly
│ │ └── ...
│ ├── dexterous manipulation
│ │ ├── data
│ │ │ ├── chunk-000
│ │ │ │ ├── episode_000000.parquet
│ │ │ │ ├── episode_000001.parquet
│ │ │ │ ├── episode_000002.parquet
│ │ │ │ ├── ...
│ │ │ ├── chunk-001
│ │ │ │ ├── ...
│ │ │ ├── ...
│ │ ├── meta
│ │ │ ├── episodes.jsonl
│ │ │ ├── episodes_stats.jsonl
│ │ │ ├── info.json
│ │ │ ├── modality.json
│ │ │ ├── stats.json
│ │ │ ├── tasks.jsonl
│ │ ├── videos
│ │ │ ├── chunk-000
│ │ │ │ ├── observation.images.front
│ │ │ │ │ ├── episode_000000.mp4
│ │ │ │ │ ├── episode_000001.mp4
│ │ │ │ │ ├── ...
│ │ │ │ ├── ...
│ │ │ ├── chunk-001
│ │ │ │ ├── ...
│ │ │ ├── ...
│ ├── pick_and_place
│ │ └── ...
│ ├── ...
├── sim
│ ├── ...
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:
{
"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
}
}
}
📥 Usage
1. Environment Setup
We recommend using conda to manage dependencies:
conda create -n dexora python=3.10 -y
conda activate dexora
# Clone this repository
git clone <your-dexora-repo-url>.git
cd Dexora
# Install Python dependencies (example)
pip install -r requirements.txt
If you plan to train or fine-tune VLAs, please ensure that PyTorch, CUDA, and any required simulation backends (e.g., Isaac, Mujoco) are installed according to your hardware.
2. Downloading the Dataset
- Simulated Dexora: Download links will be provided on the project page (see badge above).
- Real-World Dexora: High-resolution teleoperation data (RGB, proprio, actions) will be hosted via a public storage service (e.g., academic server / cloud bucket).
Typical usage:
# Example directory where you store data
export DEXORA_DATA=/path/to/dexora
# (Optional) Symlink data into this repo
ln -s $DEXORA_DATA data
3. Loading Episodes (Example)
Below is a minimal Python snippet illustrating how to load a Parquet episode from the real-world dataset:
import pandas as pd
from pathlib import Path
# Example: Loading a Parquet episode from the real-world dataset
root = Path("data/real/dexterous_manipulation/data/chunk-000")
episode_path = root / "episode_000000.parquet"
# Load trajectory using pandas
df = pd.read_parquet(episode_path)
# Access data columns (Observation, Action, Proprioception)
# Note: Columns are typically flattened in Parquet format
print("Available keys:", df.columns)
print("Actions shape:", df["action"].shape) # Example access
print("Language Instruction:", df["language_instruction"][0])
📜 Citation
If you find Dexora useful in your research, please consider citing our paper:
@misc{dexora2026,
title = {Dexora: Open-Source VLA for High-DoF Bimanual Dexterity},
author = {Dexora Team},
year = {2026},
archivePrefix = {arXiv},
eprint = {xxxx.xxxxx},
primaryClass = {cs.RO}
}
For questions, collaborations, or feedback, please feel free to open an issue or contact the maintainers via the project page.