DEXHAND-70K/dexhand-70k-review-sample

DexHand-70K: A Multi-Modal Dataset for Studying Scaling Laws in High-DOF Dexterous Manipulation

70,000 teleoperated grasping demonstrations with a 16-DOF LEAP Hand, synchronized RGB (4 cameras) + tactile forces (51-dim) + proprioception (16-dim) at 30 Hz, collected across 101 objects × 5 environments × 10 lighting conditions under a controlled factorial domain design.

Companion to the NeurIPS 2026 Evaluations & Datasets Track submission.

Why this dataset

Existing manipulation datasets either lack tactile sensing, use low-DOF parallel-jaw grippers (6–7 DOF), or have uncontrolled visual domains. DexHand-70K is the first large-scale manipulation dataset that simultaneously offers:

• High-DOF action space: 16-DOF anthropomorphic LEAP Hand
• Multi-modal sensing: synchronized vision + tactile + proprioception
• Geometric diversity: 101 objects spanning 9 functional categories, both convex and non-convex, three complexity levels
• Controlled visual domains: full factorial 5 environments × 10 lighting conditions (50 visual domains per object variant)
• Scale: 70,000 trajectories, ~6.88M frames, ~63.7 hours at 30 Hz

This design enables three research questions that no prior dataset supports:

Q	Theme	What the dataset enables
Q1	Data scaling laws in high-DOF action spaces	Power-law fits across 100 → 70K trajectories
Q2	Vision–tactile complementarity	Modality ablations at scale
Q3	Visual domain scaling	Holds data fixed, varies # training domains 5 → 100

Quick stats

Metric	Value
Trajectories	70,000
Objects	101 (280 variants)
Environments	5 (factory, hospital, living_room, office, warehouse)
Lighting conditions	10 (l1–l10)
Domain combinations	50 (full factorial)
Sampling rate	30 Hz (dt = 0.033 s)
Trajectory length	min 36 / median 76 / mean 98 / max 698 steps
Total frames	6,876,900
Total duration	~63.7 hours
Total storage	2.09 TB (gzip-compressed HDF5)
Dataset SHA-256	e3c67df11b4c5a0d6fae111542dae8ef926bfee48c8724f00b295cec5c1cbd85
License	CC-BY-4.0

Repository layout

dataset_release_70k/
├── README.md                      # this file
├── croissant.json                 # ML Croissant metadata (NeurIPS E&D Track)
├── manifest.json                  # dataset-level summary + fingerprint
├── sha256sums.txt                 # 70,000 lines: <sha256>  data/<rel>.hdf5
├── index.csv                      # per-trajectory index (relpath, object, env, light, num_steps, sha256, …)
├── index.parquet                  # same as index.csv, columnar (zstd)
├── metadata/
│   ├── objects.json               # 101 objects with taxonomy + counts
│   ├── environments.json          # 5 envs × 10 lights with descriptions + counts
│   ├── statistics.json            # length distribution + modality schema
│   └── splits/default.json        # object/env/lighting splits + scaling subsets
└── data/
    └── {object_name}/
        └── {environment}_{lighting}/
            └── demo_{variant_idx:02d}_{demo_seq:02d}.hdf5

Per-trajectory HDF5 schema

Every .hdf5 file is a single demonstration trajectory of length T steps:

Path	Shape	Dtype	Description
/rgb/cam1 … /rgb/cam4	(T, 256, 256, 3)	uint8	4 fixed-viewpoint RGB cameras (front, side, top-down, wrist-adjacent), gzip level-4 compressed
/actions	(T, 16)	float32	LEAP Hand 16-DOF target joint positions (column order a_0…a_15)
/forces	(T, 51)	float32	Tactile contact forces: 17 finger links × 3 axes (f_x, f_y, f_z)
/positions	(T, 16)	float32	Measured joint positions
/velocities	(T, 16)	float32	Measured joint velocities
/timestamps	(T,)	float64	Synchronized step timestamps in seconds (= arange(T) × 0.033)

Top-level HDF5 attributes:

Attribute	Type	Description
object_name	str	Base object identifier (one of 101)
pose_idx	int	0 if no pose-variant, otherwise the pose number (1, 2, …)
variant_idx	int	Variant index within the base object
demo_idx	int	Sequential demonstration index within (object, variant, env, light)
environment	str	One of factory / hospital / living_room / office / warehouse
lighting	str	One of l1 … l10
num_steps	int	Trajectory length (= T)
dt	float	Step duration in seconds (0.033)
description	str	Natural-language task description (e.g. "Grasp the apple object.")
robot_init_pos, robot_init_quat	(3,)/(4,) f32	Initial robot base pose
object_init_pos, object_init_quat	(3,)/(4,) f32	Initial object pose

Loading

Install the helper package:

pip install dex-hand-dataset

from dex_hand_dataset import load

# Load full dataset, lazily reads each trajectory on demand
ds = load("/path/to/dataset_release_70k", load_rgb=True)
print(len(ds))           # 70000

sample = ds[0]
sample["actions"]        # (T, 16) float32 tensor
sample["forces"]         # (T, 51) float32
sample["positions"]      # (T, 16)
sample["velocities"]     # (T, 16)
sample["rgb"]["cam1"]    # (T, 256, 256, 3) uint8
sample["meta"]           # dict: object_name, environment, lighting, …

Or read files directly:

import h5py
with h5py.File("data/apple1/factory_l1/demo_00_00.hdf5", "r") as f:
    rgb = f["rgb/cam1"][:]        # (T, 256, 256, 3) uint8
    actions = f["actions"][:]     # (T, 16) float32
    forces = f["forces"][:]       # (T, 51) float32
    print(dict(f.attrs))

Filtering with the index

Use index.parquet (or index.csv) for fast metadata-driven filtering — no need to open 70,000 HDF5 files to discover what is in the dataset:

import pandas as pd
idx = pd.read_parquet("index.parquet")
factory_l1 = idx[(idx.environment == "factory") & (idx.lighting == "l1")]
electronics = idx[idx.object_name.isin([...])]  # join with metadata/objects.json

Data collection

Demonstrations are recorded in NVIDIA Isaac Sim using a 16-DOF LEAP Hand with two teleoperation interfaces:

1. VR teleoperation — Meta Quest hand tracking → Unity retargeting → TCP → Isaac Sim @ 30 Hz
2. LEAP Hand machine — physical LEAP Hand acts as input device, joint encoders stream directly to the simulated hand

Each operator's task is to grasp the object on a platform; the platform descends and the grasp is successful if the object stays in the hand for 2 seconds.

The factorial design renders every object variant in all 50 (environment × lighting) combinations, decoupling environment geometry from lighting effects.

All 87,250 collected trajectories were audited for integrity (0 corrupted, 0 empty, 0 missing modalities); 70,000 were retained according to the quality criteria:

• each (object, variant) ≥ 5 demonstrations
• each object ≥ 2 variants (5 single-variant objects retained for category diversity: stapler1, spraybottle1, heatgun1, beaker, waterbottle)
• each kept variant contributes its 5 shortest demonstrations × all 50 (env × light) cells → 250 trajectories per variant × 280 variants = 70,000

Splits

metadata/splits/default.json defines three orthogonal evaluation axes (per the paper):

• Object generalization: 50 seen + 20 unseen objects (sampled across all 9 categories, balanced convex / non-convex)
• Environment generalization: 4 seen (warehouse, factory, hospital, living_room) + 2 unseen (bedroom, corridor) used at evaluation time
• Lighting generalization: 5 seen (l1–l5) + 3 unseen (l6, l7, l8)
• Data scaling subsets: 100, 1K, 5K, 10K, 30K, 70K

(Authoritative train/eval object lists will be finalized prior to camera-ready; users can construct custom splits from index.csv.)

Verification

The release ships a tamper-evident fingerprint and per-file SHA-256:

# Verify all 70,000 HDF5 files
sha256sum -c sha256sums.txt

# Verify dataset-level fingerprint matches manifest.json
sha256sum sha256sums.txt   # should print e3c67df1…1cbd85

The repo-level SHA-256 in croissant.json is computed as sha256(sorted concat of "<sha256> <relpath>\n" lines for every .hdf5) — equivalently, sha256(sha256sums.txt).

Object taxonomy

Each of the 101 objects is annotated along four dimensions in metadata/objects.json:

• Functional category (9): electronic, hand_tool, household, industrial, medical, measurement, office, power_tool, container
• Geometry complexity: simple / moderate / complex
• Primary material: metal / plastic / mixed / organic / glass / paper / ceramic / foam / rubber / fabric
• Size: small / medium / large

All meshes are generated via Hunyuan3D and licensed under CC-BY-4.0.

Limitations

• Simulation only. Contact physics are approximated by Isaac Sim's rigid-body solver; tactile readings are simulated contact forces, not real tactile-sensor outputs. A sim-to-real gap exists.
• Single-hand grasping. Bimanual and in-hand manipulation are not represented.
• Fixed cameras. All four cameras have fixed poses; no active or ego-centric vision.
• No depth modality. Only RGB is released (depth was never collected; see DR-0003 in the paper repository).
• Variable trajectory length. Lengths range 36–698 steps (mean 98) due to operator skill and object difficulty.

License & citation

This dataset is released under CC-BY-4.0 (Creative Commons Attribution 4.0).

Citation: To be announced with the NeurIPS 2026 publication. The paper introduces this dataset under the name DexHand-70K.

Asset attributions

• LEAP Hand URDF — dex-urdf (MIT license)
• 3D meshes — generated via Hunyuan3D (Tencent), released under CC-BY-4.0
• Simulator — NVIDIA Isaac Sim / Isaac Lab
• Teleoperation hardware — Meta Quest 3 (VR hand tracking), physical LEAP Hand