DexUMI Collection Device

Robot-free data collection device based on a vision–inertial fusion UMI architecture, 300 g+ per unit, supporting on-site data collection and training preparation.

DexUMI is the robot-free data collection device in the LingYu TA robot ecosystem. Designed for TA robots, the data it collects can be used for TA model training and migrated to real-robot validation, reducing the cost of moving from data collection to physical deployment. The device tracks the left and right hands independently, recording posture changes and operation trajectories in real time; each unit weighs 300 g+ and supports offloading compute to a backpack or external computing unit, making it suitable for portable data collection and training preparation.

DexUMI collection device product photo

Core Advantages

TA Isomorphism

Designed for TA robots; collected data can be used for model training and real-robot validation.

Lightweight Design

300 g+ per unit, supporting up to 1 kg of applied force, suitable for portable data collection.

High-precision Tracking

Independent left- and right-hand tracking, recording posture changes and operation trajectories in real time.

Multimodal Perception

Captures RGB video, depth, IMU and audio data with hardware-level timestamp alignment.

Rapid Deployment

No mapping required; supports data collection and training preparation across a range of scenarios.

Technical Parameters

VR Headset

Head Binocular Resolution1920(H) × 1920(V) × 2(View)
Frame Rate45 FPS
Field of ViewHFOV 120° / VFOV 120°
LatencyGlass-to-Glass < 80 ms
Weight680 g (including head binocular)

UMI Gripper

Wrist Binocular Resolution1280(H) × 800(V) × 2(View)
Frame Rate45 FPS
Field of ViewHFOV 120° / VFOV 76°
Gripper TypeTwo-finger
Max Opening Width100 mm
Max Gripping Weight1 kg
Dimensions166 × 154 × 131.5 mm
Weight300 g+ (single)

Compute Unit

CPU6 × ARM Cortex-A78AE
BPU1 × BPU Nash, 80 TOPS
MCU4 × ARM Cortex-R52
Storage64 GB eMMC + 1 TB SSD
Memory12 GB LPDDR5
LANGigabit Ethernet × 2
OSUbuntu
Weight540 g

Application Scenarios

Human-Robot Alignment & Imitation Learning

Build robot training demonstration datasets.

Fine Manipulation Learning

Supports tasks such as grasping, assembly and tool use.

Long-sequence Task Modeling

Supports continuous tasks such as pouring, tidying and packing.

CONTACT

Contact Us

Product inquiries, data services, partnerships and press all start here. Tell us what you need in the form.

Fields marked * are required

China mainland 11-digit mobile, or international format with + (e.g. +85212345678)

0/8000

Your details are used for this inquiry only.

Cookies on this site

We use analytics cookies to understand how this site is used so we can improve it. They are only enabled with your consent, and declining does not affect any part of the site. Read our Privacy Policy