WANDA: Worlds in One Demo

A Synthetic Data Engine for Learning Open-World Mobile Manipulation

🌐 Project page: https://wanda.lecar-lab.org/ · 📄 Paper (PDF) · 💻 Code (coming soon) · 🕹️ Interactive 4D viewer

Lingxiao Guo*, Huanyu Li*, Guanya Shi — Carnegie Mellon University _{*Equal contribution; order decided by a coin flip.}

---

WANDA is a synthetic data engine that turns one human demonstration into diverse training data for open-world mobile manipulation. From a single RGB-D demo, WANDA reconstructs the scene (Gaussian-splat background, object meshes, and interaction trajectories), then re-generates it across new spatial configurations, scenes, and even robot embodiments — producing the large-scale demonstration sets used to train the vision-language-action (VLA) policies in the paper.

Robot data has been priced in human hours. WANDA prices it in compute.

This repository hosts the WANDA-generated training datasets for the five long-horizon tasks evaluated in the paper, each learned from a single real human demonstration on the Agibot G1. Policies trained on this data reach 54.8% task progress on real-world rollouts and outperform policies trained on 50 teleoperated demonstrations — while one hour of training data is generated in ≈3 GPU-hours.

Dataset overview

Subset	Task instruction	Episodes	Task chunks
drop_trash	Pick up the empty can from the table and drop it into the trash bin.	3,496	task-0051 … 0054
fridge	Pick up the snack bag, put it into the fridge, and close the door.	436	task-0056
lunch_box	Close the lunch box on the table, pick it up, and throw it into the sink.	2,043	task-0052 … 0054
pour	Pick up the kettle and pour water into the cups.	7,775	task-0055 … 0062
utensils	Pick up the utensils from the table, place them in the basket, and put the basket in the sink.	3,172	task-0053 … 0056
Total		16,922	~26.8M timesteps (≈248 h @ 30 FPS)

Each subset is a self-contained LeRobot v2.1 dataset (meta/, data/, videos/). Different task chunks within a subset correspond to different spatial configurations / scenes generated by WANDA for the same task.

Format

• Robot: agibot_g1 (dual-arm mobile manipulator) · Control/record rate: 30 FPS · Codebase: LeRobot v2.1

Per-frame features (parquet):

Key	Dtype	Shape	Description
observation.state	float32	[20]	Proprioceptive robot state
action	float32	[20]	Action target
observation.cam_rel_poses	float32	[21]	Camera poses relative to the robot
observation.task_info	float32	[46]	Task/goal conditioning vector
observation.images.rgb.head	video	[800, 1280, 3]†	Head camera
observation.images.rgb.left_wrist	video	[480, 848, 3]	Left-wrist camera
observation.images.rgb.right_wrist	video	[480, 848, 3]	Right-wrist camera
timestamp, index, episode_index, task_index	—	—	LeRobot bookkeeping

_{† Head-camera resolution is 800×1280 for all subsets except utensils (480×768).}

Directory layout (per subset):

<subset>/
├── meta/
│   ├── info.json            # features, fps, robot_type, totals
│   ├── tasks.jsonl          # natural-language task instructions
│   ├── episodes.jsonl       # per-episode index
│   └── episodes_stats.jsonl # per-episode feature statistics
├── data/
│   └── task-XXXX/episode_XXXXXXXX.parquet
└── videos/
    └── task-XXXX/observation.images.rgb.<cam>/episode_XXXXXXXX.mp4

Usage

Download a single subset and load it with LeRobot:

from huggingface_hub import snapshot_download
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

root = snapshot_download(
    "LeCAR-Lab/Wanda",
    repo_type="dataset",
    allow_patterns="utensils/*",   # one of: drop_trash, fridge, lunch_box, pour, utensils
)
ds = LeRobotDataset(repo_id="LeCAR-Lab/Wanda", root=f"{root}/utensils")
print(ds)
print(ds[0].keys())

Or read the raw parquet directly (no LeRobot required):

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "LeCAR-Lab/Wanda",
    "utensils/data/task-0053/episode_00530010.parquet",
    repo_type="dataset",
)
df = pd.read_parquet(path)
print(df.shape, df.columns.tolist())

License

Released under CC BY 4.0. If you use this dataset, please provide attribution and cite the paper below.

Citation

@article{guo2026wanda,
  title   = {Worlds in One Demo: A Synthetic Data Engine for Learning Open-World Mobile Manipulation},
  author  = {Guo, Lingxiao and Li, Huanyu and Shi, Guanya},
  year    = {2026}
}

Links

• Project page & interactive 4D viewer: https://wanda.lecar-lab.org/
• Paper (PDF): https://wanda.lecar-lab.org/assets/paper/WANDA.pdf
• Code: coming soon