OOJU/wild-mani-kitchen

WILD-Mani: A Real-World Bimanual Manipulation Dataset – Kitchen Edition

The first installment of the WILD-Mani series.

• High-fidelity bimanual demonstrations collected in real-world kitchen environments using Mixed Reality (MR).
• Designed to support Sim2Real transfer for bimanual manipulation tasks.
• Current Domain: Kitchen

Update: 02-21-2026

We are currently scaling the WILD-Mani series to 1,000+ bimanual episodes with full Unitree G1 kinematic validation. Access to the expanded library and our collection/automated-validation pipeline is being released in cohorts.

Dataset Description

This dataset contains human demonstrations of kitchen manipulation tasks captured using a Mixed Reality headset in real-world kitchen environments. The data tracks both hands simultaneously with 6-DOF pose tracking, enabling bimanual robot imitation learning research.

Key Features

• Real-world environments: Demonstrations captured in actual kitchens (not simulation)
• Mixed Reality capture: High-fidelity hand tracking in physical spaces
• Bimanual manipulation: Both hands tracked simultaneously
• Camera view: Single camera (center eye) per demonstration
• Diverse tasks: Multiple kitchen manipulation tasks with varying complexity

Dataset Statistics

Property	Value
Episodes	102
Total Frames	79,244
FPS	30
Cameras	1
Environment	Real kitchen with lighting variations

Task Categories

Category	Task	Objects	Actions	Demos
Meal Prep	Get item from fridge	Bottle, container	Open, pick, close, place	25
Dish Organizing	Place cups on shelf	Cups (2-3)	Pick, place	26
Dish Organizing	Put utensils in drawer	Fork, spoon, knife	Pick, open, place, close	26
Meal Prep	Set table (plate + utensil)	Plate, fork, knife	Pick, place	25

Start/End Labeling (Temporal Bounds)

Each episode has start and end labels so that the stored trajectory contains only the task-relevant segment. Frames before the task start (e.g. approach to the scene or button press) and after the task end are excluded. Users can assume that the data does not include unnecessary motion at the beginning or end of each demonstration; each episode is trimmed to the intended task interval.

Action Labels (observation.action_label)

Per-frame labels are provided as action::subaction strings. The vocabulary varies by task. Below is the label set for each task.

• Hierarchical Semantic Labeling (VLA-Ready)
• Language-Conditioned Policy training and Chain-of-Thought reasoning for bimanual manipulation.

Get item from fridge (25 episodes)

Action	Subactions
open	approach → grasp → pull → release
pick	reach → grasp → extract
close	approach → push → retract
place	transport → release

Labels: open::approach, open::grasp, open::pull, open::release, pick::reach, pick::grasp, pick::extract, close::approach, close::push, close::retract, place::transport, place::release.

Place cups on shelf (26 episodes)

Action	Subactions
pick	approach → align → grasp → lift
place	release → retract

Labels: pick::approach, pick::align, pick::grasp, pick::lift, place::release, place::retract.

Put utensils in drawer (26 episodes)

Action	Subactions
open	approach → grasp → pull → release
pick	approach → grasp → lift
place	transport → release
close	approach → push → retract

Labels: open::approach, open::grasp, open::pull, open::release, pick::approach, pick::grasp, pick::lift, place::transport, place::release, close::approach, close::push, close::retract.

Set table (plate + utensil) (25 episodes)

Action	Subactions
pick	approach → align → grasp → lift
place	transport → adjust → lower → release

Labels: pick::approach, pick::align, pick::grasp, pick::lift, place::transport, place::adjust, place::lower, place::release.

---

Diversity Dimensions

Dimension	Coverage	Details
Actions	4+ types	Pick, place, open, close
Objects	15+ objects	Kitchen items varying in size, shape, material
Environment	1 kitchen	Natural light, artificial light, dim lighting
Surfaces	3 surfaces	Counter, dining table, shelf
Task Complexity	Single + Multi-step	Pick-place (single) → Fridge sequence (multi-step)
Clutter	Natural	Objects on counter, objects in fridge

Data Format

This dataset follows the LeRobot v3.0 format.

Action Space (14D)

The action space represents delta poses for both hands:

Dimensions	Description
0-2	Left hand position delta (x, y, z)
3-6	Left hand rotation delta (quaternion x, y, z, w)
7-9	Right hand position delta (x, y, z)
10-13	Right hand rotation delta (quaternion x, y, z, w)

Observation Space

Key	Shape	Description
state	(14,)	Combined left/right wrist poses
left_state	(7,)	Left wrist pose (position + quaternion)
right_state	(7,)	Right wrist pose (position + quaternion)
right_joint_poses	(78,)	Right hand full skeleton (26 joints × 3D position)
left_joint_poses	(78,)	Left hand full skeleton (26 joints × 3D position)
observation.images.center_eye	(H, W, 3)	Center eye camera view
observation.action_label	string	Per-frame action label (e.g. pick::grasp, place::release)

Joint Order (26 joints per hand)

palm, wrist, thumb_metacarpal, thumb_proximal, thumb_distal, thumb_tip, index_metacarpal, index_proximal, index_intermediate, index_distal, index_tip, middle_metacarpal, middle_proximal, middle_intermediate, middle_distal, middle_tip, ring_metacarpal, ring_proximal, ring_intermediate, ring_distal, ring_tip, little_metacarpal, little_proximal, little_intermediate, little_distal, little_tip

Recording Setup

• Headset: Meta Quest 3 (Mixed Reality)
• Environment: Real-world kitchen
• Tracking: Full skeletal hand tracking (26 joints per hand) via Quest 3 hand tracking
• Cameras: 1 camera (center eye view)
• Software: Unity-based recording system

File Structure

.
├── data/
│   ├── chunk_0000.parquet
│   └── chunk_0001.parquet
├── meta/
│   ├── info.json
│   ├── stats.json
│   └── episode_index.parquet
├── videos/
│   └── observation.images.center_eye/
└── README.md

License

This dataset is released under the Apache 2.0 License.

Citation

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

@misc{wild_mani_kitchen,
  title={WILD-Mani: A Real-World Bimanual Manipulation Dataset -- Kitchen Edition},
  author={OOJU},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/OOJU/wild-mani-kitchen}}
}

Acknowledgments

Dataset collected using Mixed Reality hand tracking in real-world kitchen environments. Recording system built with Unity and Meta Quest.