1.05k episodes · 30 fps · 6 cameras · 800×600 h264
SenseXperience UMI Human Demonstration Dataset (LeRobot)
This open dataset showcases the real-world data acquisition capability of SenseXperience, our wearable embodied AI data collection platform.
It provides 1,051 episodes, 220,963 frames, and 92 manipulation tasks collected from natural human demonstrations with synchronized multimodal sensors.
Interactive Dataset Browser:
- UMI Dataset: https://io-ai.tech/lerobot/?url=sample%3A%2F%2Fsensexperience_umi
- EGO Dataset: https://io-ai.tech/lerobot/?url=sample%3A%2F%2Fsensexperience_ego
Dataset Characteristics
- Human demonstrations are collected with a wearable first-person setup.
- No master-arm teleoperation is used during data collection.
- Data is recorded from synchronized multimodal sensors in real environments.
- The released format is compatible with LeRobot workflows.
Quick Facts
| Item | Value |
|---|---|
| Collection Platform | SenseXperience UMI |
| Dataset Version | v3.0 |
| Episodes | 1,051 |
| Frames | 220,963 |
| Tasks | 92 |
| Cameras | 6 synchronized views |
| FPS | 30 |
| Storage | Parquet + MP4 (H.264) |
What Makes SenseXperience Different
Compared with traditional teleoperation pipelines, SenseXperience emphasizes:
- Natural human demonstrations in real scenarios.
- Lower hardware and deployment complexity.
- Better scalability for large embodied datasets.
- A direct path from collection to LeRobot training.
Tasks
The dataset includes 92 real-world manipulation tasks, such as:
- Pick-and-place and object transfer
- Shelf and container interaction
- Household interaction
- Dexterous bimanual manipulation
- Human-object interaction
Data Schema and Sensors
Camera Observations
| Camera | Resolution | Role |
|---|---|---|
camera_head_left |
2048x1536 | Left egocentric head view |
camera_head_right |
2048x1536 | Right egocentric head view |
camera_left_wrist |
2048x1536 | Left wrist manipulation view |
camera_right_wrist |
2048x1536 | Right wrist manipulation view |
camera_depth_head |
800x600 | Head depth camera (RGB-encoded) |
camera_gripper |
800x600 | Gripper camera |
All videos are time-synchronized and encoded with H.264 at 30 FPS.
Human Motion Features
7-DoF poses are provided for:
- Head
- Left gripper
- Right gripper
Pose format:
[x, y, z, qx, qy, qz, qw]
Gripper state format:
[
left_Left_Gripper_IntegratedPose_rad,
left_Left_Gripper_IntegratedPose_dist,
right_Right_Gripper_IntegratedPose_rad,
right_Right_Gripper_IntegratedPose_dist
]
Feature Specification
| Feature | Type | Shape |
|---|---|---|
observation.images.camera_head_left |
video | (1536, 2048, 3) |
observation.images.camera_head_right |
video | (1536, 2048, 3) |
observation.images.camera_left_wrist |
video | (1536, 2048, 3) |
observation.images.camera_right_wrist |
video | (1536, 2048, 3) |
observation.images.camera_depth_head |
video | (600, 800, 3) |
observation.images.camera_gripper |
video | (600, 800, 3) |
observation.pose.head |
float32 | (7,) |
observation.pose.left_gripper |
float32 | (7,) |
observation.pose.right_gripper |
float32 | (7,) |
observation.state |
float64 | (4,) |
Dataset Layout
data/
chunk-000/
file-000.parquet
videos/
observation.images.camera_head_left/
chunk-000/
file-000.mp4
Quick Start (LeRobot)
from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("ICRA-WBCD/IO-AI-SenseXperience-UMI")
Recommended Uses
- Embodied AI pretraining
- Vision-Language-Action (VLA) training
- Imitation learning from human demonstrations
- Human-to-robot transfer research
- Multiview manipulation understanding
Limitations and Roadmap
Current Limitations
- Mainly focused on indoor manipulation tasks.
- Human demonstrations include natural execution variance.
- Depth observations are currently RGB-encoded videos.
- This release does not yet include robot action labels.
Planned Extensions
- Robot action annotations
- Hand keypoints
- More environments and task categories
- Language annotations
- Calibration metadata
Learn More About SenseXperience
This dataset is an open demonstration of SenseXperience data collection quality and scalability.
- Product website: https://io-ai.tech/en/sensexperience/
- Technical specs: https://io-ai.tech/en/sensexperience/specs/
- Company website: https://io-ai.tech/en/
- Contact for pilot projects or custom data programs: io@io-ai.tech
License
Released under the Apache-2.0 License.
Citation
@dataset{sensexperience_umi_dataset,
title={SenseXperience UMI Human Demonstration Dataset},
author={IO-AI.tech},
year={2026},
publisher={io <io@io-ai.tech>}
}
- Downloads last month
- 630