willi19/autodex-gallery

AutoDex: Real-World Dexterous Grasp Experience

Autonomously building reusable grasp knowledge through real-world trials.
CoRL 2026

Project · Paper · Interactive Gallery · Code

AutoDex lets a robot try grasps, verify outcomes, and reset objects autonomously—turning real-world experience into a reusable dexterous grasp library. Browse recorded trials in the interactive gallery, or download grasp annotations, robot trajectories, and multi-view recordings for your own experiments.

This release contains 2,610 real-robot grasp trials with Allegro and Inspire hands, with synchronized multi-view video, calibrated cameras, 6-DoF object poses, and executed grasp annotations.

Subset	Robot	Trials
allegro	xArm6 + Allegro (16-DoF)	2,226
inspire	xArm6 + Inspire (6-DoF)	384

What is included?

Data	Contents
Executed grasps	Wrist poses in the object frame and hand joint configurations
Robot motion	Arm and hand trajectories
Multi-view observations	Camera videos, intrinsics, and extrinsics
Object geometry in the scene	Object poses and, where available, tracked object trajectories
Trial metadata	Grasp timing, execution stages, and success annotations
Interactive previews	Animated 3D assets, thumbnails, and gallery videos

Executed wrist poses are expressed in the object frame, allowing grasps to be reused relative to the object.

Start with one object

Install the download client:

pip install -U huggingface_hub numpy

Download the Allegro trials for one object:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="willi19/autodex-gallery",
    repo_type="dataset",
    local_dir="autodex",
    allow_patterns=["allegro/attached_container/*"],
)

To download both hand subsets, replace allow_patterns with:

allow_patterns=["allegro/*", "inspire/*"]

These patterns select the trial data without downloading the separate gallery asset folders.

Browser-friendly videos

Compressed MP4 videos used by the interactive gallery are hosted in the companion autodex-gallery-video repository. Download the corresponding object with:

snapshot_download(
    repo_id="willi19/autodex-gallery-video",
    repo_type="dataset",
    local_dir="autodex-videos",
    allow_patterns=["allegro/attached_container/*"],
)

File organization

Trial data is organized by hand, object, and recording timestamp:

{hand}/{object}/{timestamp}/
├── pose_world.npy          # Object-to-world transform
├── R2C.npy                 # Robot-base-to-world transform
├── cam_param/              # Camera intrinsics and extrinsics
├── videos/                 # Camera recordings
├── arm/                    # Arm trajectories
├── hand/                   # Hand trajectories
├── executed_grasp/
│   ├── wrist_se3.npy       # Wrist pose in the object frame
│   ├── grasp_pose.npy      # Hand configuration at grasp
│   └── meta.json           # Timing, execution stages, and annotations
├── object_tracking/       # Object tracking outputs, where available
└── recompute_pose.json    # Pose-quality metadata

Additional top-level folders (interactive_3d/, turntable/, view_thumbnails/, experiments/, and experiments_proxy/) support the gallery. Available files and annotations can vary between trials.

Load a grasp

import json
from pathlib import Path

import numpy as np

trial = Path("autodex/allegro/attached_container/20260121_163413")

object_to_world = np.load(trial / "pose_world.npy")
robot_to_world = np.load(trial / "R2C.npy")
wrist_in_object = np.load(trial / "executed_grasp/wrist_se3.npy")
finger_joints = np.load(trial / "executed_grasp/grasp_pose.npy")
with (trial / "executed_grasp/meta.json").open() as f:
    metadata = json.load(f)

# Transform the recorded wrist pose into the robot-base frame.
wrist_in_robot = np.linalg.inv(robot_to_world) @ object_to_world @ wrist_in_object

print("Hand configuration:", finger_joints)
print("Trial metadata:", metadata)

The transforms above are 4×4 homogeneous matrices. For a new placement, use the object's current pose and the calibration for your setup rather than the recorded scene pose.

Citation

@misc{choi2026autodex,
  title = {AutoDex: An Automated Real-World System for Dexterous Grasping Data Collection},
  author = {Choi, Mingi and Kim, Gunhee and Kim, Jisoo and Kim, Taeksoo and
            Ha, Taeyun and Lim, Jongbin and Joo, Hanbyul},
  year = {2026},
  eprint = {2606.23689},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO}
}

Mingi Choi, Gunhee Kim, Jisoo Kim, Taeksoo Kim, Taeyun Ha, Jongbin Lim, and Hanbyul Joo · SNU and RLWRLD

For questions about the data, open a discussion on this dataset repository or contact [email protected].