🤖 PalmDex: An Embodiment-Agnostic Tactile Manipulation Dataset

A growing, multimodal robotic manipulation dataset featuring synchronized dual-camera video, dual-hand tactile sensing, and hand pose tracking across diverse real-world environments. All demonstrations are collected via human teleoperation — without any specific robot embodiment — and annotated at the action-segment level with rich categorical labels.

This is a living dataset. New environments, tasks, and objects will be added over time. The current release covers chemistry laboratory operations. Future releases will expand to additional scenes such as kitchens, workshops, offices, grocery stores, and more. See the Annotation Schema for the full list of supported environments and object categories.

📋 Dataset Summary

Property	Value
Design philosophy	Embodiment-agnostic — human hand demonstrations with tactile & visual sensing, transferable to any robot morphology
Environments	Multi-scene (currently: chemistry lab; planned: kitchen, workshop, office, store, etc.)
Current tasks	Mortar & pestle grinding, Crucible tongs transfer
Total annotated segments	1,001 (and growing)
Total raw episodes	~300+
Modalities	Ego-view video, side-view video, dual-hand tactile pressure (256-d × 2), hand pose (72-d)
Video resolution	640 × 480 @ 30 FPS
Video codec	AV1 (raw), H.264/MP4 (annotated segments)
Tactile sensor	Dual tactile gloves, 256 channels per hand, float32
Total size	~8.6 GB (current release)
Data format	LeRobot v3.0 (raw) + JSONL annotations + Apache Parquet (tactile)

---

💡 Why Embodiment-Agnostic?

Traditional robot manipulation datasets are tightly coupled to a specific robot platform. This dataset takes a fundamentally different approach:

• Human hands as the universal demonstrator — Demonstrations are performed by human operators wearing sensorized tactile gloves, capturing the natural dexterity, force modulation, and multi-finger coordination that is difficult to obtain from any single robot embodiment.
• Transferable to any morphology — The rich multi-modal signals (vision + tactile + hand pose) can be retargeted to different robot hands, grippers, or end-effectors via learned or analytic mappings.
• Scalable across environments — The portable data collection rig allows rapid deployment in diverse real-world settings without robot-specific setup, enabling the dataset to grow across many environments and tasks.

---

🗂️ Available Sub-Datasets

The dataset is organized by recording session. Each sub-dataset is a self-contained package with raw data and cleaned annotations. Sub-datasets are grouped below by environment and task.

🧪 Chemistry Lab

Task 1: Mortar and Pestle Grinding

Sub-Dataset	Segments	Action Sequence	Size
0508_1_mortar_pestle	214	picking up → grinding → placing	1.5 GB
0508_2_mortar_pestle	201	picking up → grinding → placing	2.4 GB
Subtotal	415		3.9 GB

Task 2: Crucible Tongs Transfer

Sub-Dataset	Segments	Action Sequence	Size
0508_3_crucible_tongs_transfer	349	picking up → clamping & transferring → placing → restoring	1.5 GB
0508_4_crucible_tongs_transfer	237	picking up → clamping & transferring → placing	3.2 GB
Subtotal	586		4.7 GB

🔜 Coming Soon

Additional environments and tasks are in preparation, including but not limited to:

Environment	Example Tasks	Status
Kitchen	Object sorting, tool use	Planned
Workshop	Assembly, tool manipulation	Planned
Office	Stationery handling, organizing	Planned
Grocery Store	Product picking, shelf stocking	Planned
Living Room	Everyday object manipulation	Planned

Each new environment will follow the same data format and annotation schema, ensuring full compatibility across the entire dataset.

---

🏷️ Annotation Schema

The annotation schema is designed to be environment-agnostic and extensible. New environments, objects, and actions can be added without breaking existing data.

Each annotated segment is labeled with the following multi-dimensional categorical attributes:

Label Categories

Category	Type	Description
environment	single-select	Scene / location of the task
object	single-select	Object(s) being manipulated
action	single-select	Fine-grained action label
grasp_type	single-select	Grasp taxonomy classification
notes	free-text	Annotator notes per segment

Supported Environments (Expandable)

The schema currently supports the following environments, with new entries added as data collection expands:

hardware store · office 1 · office 2 · grocery store 1 · grocery store 2
workshop · kitchen · sports store · bedroom · plant store
restaurant 1 · restaurant 2 · living room · biochemistry lab
wet lab bench · solution preparation station · distillation station
fume hood · sterile workbench · analytical instrument station
cold storage area · waste disposal area · chemistry lab · ...

Supported Grasp Types

Small Diameter · Tip Pinch · Prismatic 2 Finger · Prismatic 3 Finger
Index Finger Extension · Medium Wrap · Palmar · Fixed Hook
Light Tool · Precision Sphere · ...

Current Action Labels by Task

Mortar & Pestle Grinding:

picking up  →  grinding  →  placing

Crucible Tongs Transfer:

picking up  →  clamping and transferring  →  placing  [→ restoring object position]

Action Distribution (Current Release)

Action	0508_1	0508_2	0508_3	0508_4	Total
picking up	72	67	116	79	334
grinding	72	67	—	—	139
placing	70	67	116	79	332
clamping and transferring	—	—	116	79	195
restoring object position	—	—	1	—	1

---

📁 Directory Structure

Each sub-dataset follows a consistent structure:

<dataset_root>/
├── README.md                              ← This file (top-level)
├── <session_id>_<task_name>/              ← Sub-dataset package
│   ├── README.md                          ← Sub-dataset description
│   ├── dataset_metadata.json              ← Machine-readable metadata
│   ├── checksums.sha256                   ← Package-level integrity checksums
│   ├── raw/
│   │   └── lerobot/<raw_id>/             ← Original LeRobot v3.0 dataset
│   │       ├── data/                      ←   Parquet data files (tactile, state, etc.)
│   │       ├── videos/                    ←   Full-episode AV1 video streams
│   │       ├── images/                    ←   (if applicable)
│   │       ├── meta/                      ←   info.json, stats.json, tasks.parquet, ...
│   │       └── sync_diagnostics/          ←   Timing / synchronization logs
│   └── annotated/
│       ├── annotations.jsonl              ← All segment annotations (1 JSON per line)
│       ├── schema.json                    ← Annotation schema definition
│       ├── manifest.json                  ← File manifest for all segments
│       ├── validation_report.json         ← Automated QA report
│       ├── release_report.json            ← Release summary
│       ├── checksums.sha256               ← Annotated-subset checksums
│       └── segments/
│           └── train/
│               └── rbt_epXXXXXX_sXXXXXX/ ← Per-segment directory
│                   ├── ego.mp4            ← Ego-view video clip
│                   ├── side.mp4           ← Side-view video clip
│                   ├── data.parquet       ← Tactile + state data
│                   └── meta.json          ← Segment metadata & labels
├── ...                                    ← More sub-datasets

---

🎯 Modalities in Detail

1. Visual — Dual-Camera Video

Each episode is recorded from two synchronized camera viewpoints:

Camera	Key	Resolution	FPS	Codec
Ego-view	observation.images.ego	640 × 480	30	AV1 (raw) / H.264 (annotated)
Side-view	observation.images.side	640 × 480	30	AV1 (raw) / H.264 (annotated)

• Raw episodes: Full uncut videos stored under raw/lerobot/*/videos/.
• Annotated segments: Trimmed video clips per action segment stored as ego.mp4 and side.mp4.

2. Tactile — Dual-Hand Pressure Arrays

High-resolution tactile data from sensorized gloves, recorded at 30 Hz:

Channel	Key	Shape	Dtype	Description
Left hand	observation.tactile_left	(256,)	float32	Raw pressure values from left-hand tactile glove
Right hand	observation.tactile_right	(256,)	float32	Raw pressure values from right-hand tactile glove

• Tactile routing protocol: Header (4 bytes) + Sequence (1 byte) + Type (1 byte) + Data
• Type 0x01 → Left hand, Type 0x02 → Right hand
• Per-segment tactile data is stored in data.parquet files

3. Hand Pose — Teleoperation State

Full articulated hand pose captured via the UDexReal hand-tracking system:

Key	Shape	Dtype	Description
observation.udexreal	(72,)	float32	Dual-hand bone positions, rotations, scales, and finger joint parameters

The 72-dimensional vector encodes:

• Head bone: Location (3D), Parent, Rotation (3D), Scale (3D) — 10 dims
• Left hand: Calibration status + 24 joint parameters + joystick/button states — 31 dims
• Right hand: Calibration status + 24 joint parameters + joystick/button states — 31 dims

---

🔖 Segment ID Convention

All segment IDs follow the format:

rbt_ep{EPISODE:06d}_s{SEGMENT:06d}

• ep = Episode index (from the original LeRobot dataset)
• s = Globally unique segment index within the sub-dataset

Example: rbt_ep000002_s000001 = Episode 2, Segment 1

---

📊 JSONL Annotation Format

Each line in annotations.jsonl is a self-contained JSON object:

{
  "segment_id": "rbt_ep000002_s000001",
  "episode_idx": 2,
  "start_frame": 6,
  "end_frame": 94,
  "labels": {
    "environment": "chemistry lab",
    "object": "mortar and pestle",
    "action": "picking up",
    "grasp_type": "Small Diameter"
  },
  "outcome": {
    "success_frame": null,
    "final_success": null,
    "events": []
  },
  "notes": "研磨物体",
  "workflow_status": "approved"
}

Per-Segment meta.json Format

{
  "segment_id": "rbt_ep000002_s000001",
  "source": {
    "dataset_id": "0508",
    "episode_idx": 2,
    "start_frame": 6,
    "end_frame": 94,
    "start_time_sec": 0.2,
    "end_time_sec": 3.133,
    "fps": 30.0
  },
  "labels": {
    "environment": "chemistry lab",
    "object": "mortar and pestle",
    "action": "picking up",
    "grasp_type": "Small Diameter"
  },
  "outcome": { ... },
  "notes": "研磨物体"
}

---

🚀 Quick Start

Loading Annotations

import json

annotations = []
with open("0508_1_mortar_pestle/annotated/annotations.jsonl") as f:
    for line in f:
        annotations.append(json.loads(line))

# Filter by action
grinding_segments = [a for a in annotations if a["labels"]["action"] == "grinding"]
print(f"Grinding segments: {len(grinding_segments)}")

Loading Tactile Data

import pandas as pd

# Load per-segment tactile data
df = pd.read_parquet(
    "0508_1_mortar_pestle/annotated/segments/train/rbt_ep000002_s000001/data.parquet"
)
print(df.columns.tolist())
print(df.shape)

Loading Segment Video

import cv2

cap = cv2.VideoCapture(
    "0508_1_mortar_pestle/annotated/segments/train/rbt_ep000002_s000001/ego.mp4"
)
fps = cap.get(cv2.CAP_PROP_FPS)
frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
print(f"FPS: {fps}, Frames: {frame_count}")
cap.release()

Using with LeRobot (Raw Data)

The raw/lerobot/ directories are compatible with the LeRobot framework (codebase v3.0):

# Raw dataset can be loaded with LeRobot's dataset loader
# Refer to https://github.com/huggingface/lerobot for details

---

✅ Data Quality & Validation

• All annotated segments have passed automated validation (0 P0 critical errors)
• Validation issues are limited to P1 informational notes (P1_LEGACY_APPROVED_NEEDS_REVIEW — indicating segments migrated from a legacy annotation workflow)
• SHA-256 checksums are provided at both the package level (checksums.sha256) and the annotated subset level (annotated/checksums.sha256)

Validation Summary (Current Release)

Sub-Dataset	P0 (Critical)	P1 (Info)	P2 (Warning)
0508_1_mortar_pestle	0	214	0
0508_2_mortar_pestle	0	201	0
0508_3_crucible_tongs_transfer	0	345	0
0508_4_crucible_tongs_transfer	0	237	0

Integrity Verification

# Verify full package integrity
cd 0508_1_mortar_pestle
sha256sum -c checksums.sha256

# Verify annotated subset only
cd annotated
sha256sum -c checksums.sha256

---

📝 Processing Notes

Label Normalization

• In sub-datasets 0508_1 and 0508_4, the original moving and picking up stages were merged into a single picking up label to ensure consistency across the dataset.
• In 0508_4, the original adjusting stage was normalized to clamping and transferring.

Scope of 0508_4

The original 0508_4 LeRobot recording contains more than one task. Episodes 001–096 correspond to the crucible-tongs transfer task, while episodes starting from 097 correspond to stirring a beaker solution with a glass rod. Only the crucible-tongs transfer subset (79 cleaned episodes) is included in the annotated release. Episodes 080 and 095 were excluded during quality control.

---

🏗️ Data Collection Setup

Component	Specification
Teleoperation system	UDexReal hand-tracking gloves with integrated tactile sensing
Cameras	Dual RGB cameras (ego-view + side-view), 640×480, 30 FPS
Tactile sensors	Dual tactile gloves, 256 pressure channels per hand
Recording framework	LeRobot v3.0
Annotation tool	Custom JSONL-based annotation pipeline with schema validation
Data synchronization	Hardware-triggered, verified via sync_diagnostics/

The entire collection rig is portable and can be deployed across different environments without any robot-specific hardware, enabling rapid scaling to new scenes and tasks.

---

---

📖 Citation

If you use this dataset in your research, please cite:

@misc{tactile_manipulation_2026,
  title     = {Embodiment-Agnostic Tactile Manipulation Dataset},
  author    = {Rimbot},
  year      = {2026},
  publisher = {Hugging Face},
  note      = {A growing multimodal dataset with synchronized dual-camera
               video, dual-hand tactile sensing, and hand pose tracking
               for dexterous manipulation across diverse real-world environments}
}

---

📬 Contact

For questions, issues, or collaboration inquiries, please open an issue on this Hugging Face dataset repository.

---

📅 Changelog

Date	Update
2026-05-23	Initial release — Chemistry Lab: mortar & pestle grinding (×2) + crucible tongs transfer (×2)