The Glove That Teaches Robots How Human Hands Move
A 210-gram glove with 23 points of movement captures every motion, touch, and gesture a human hand makes and turns it into structured training data for robots.
Photo source:
Genrobot
Why DAS Dex Exists
Robots are getting smarter, but their hands are
not keeping up. Most can grip, lift, and place objects, yet they struggle with
tasks that require the kind of fine control humans take for granted, like
turning a key, folding fabric, or peeling tape from a surface. The reason is
not mechanical. Robotic hands with enough joints and sensors to mimic human
fingers already exist. The bottleneck is data. To teach a robot how to perform
a delicate task, developers need precise recordings of how a human hand performs
it first, down to the exact angle of each finger, the pressure applied at each
fingertip, and the trajectory of every movement.
GenRobot AI, a Beijing-based company founded in
July of last year, built the DAS Dex as a robot training data glove designed to
close that gap. Instead of relying on cameras watching hands from the outside
or expensive teleoperation rigs that cost tens of thousands of dollars, DAS Dex
is a lightweight wearable that captures human hand behavior from the inside. A
person puts it on in five seconds, performs everyday tasks naturally, and the
glove records everything needed to teach a robot to replicate those same
movements.
How the Robot Training Data Glove Captures Movement
DAS Dex tracks 23 degrees of freedom across the
hand, matching the number of independent movements a human hand can perform.
Each joint is measured by a 3-millimeter magnetic encoder with a precision of
0.02 degrees. Fingertip position is accurate to within one millimeter. The
entire glove weighs 210 grams, roughly the weight of a smartphone, and is
designed to feel natural enough that the wearer can perform tasks without the
device interfering with their normal hand behavior.
Beyond motion, the glove captures tactile data.
Sensors distributed across the fingers detect forces as light as 0.05 Newtons
with spatial resolution at the millimeter scale in three dimensions. This means
the system records not only where each finger went, but how much pressure it
applied and where on the fingertip that pressure was concentrated. A
wrist-mounted camera with a 150-degree field of view captures a full visual
record of each task from the hand's own perspective. All of this data-
fingertip trajectories, joint angles, tactile maps, spatial positioning, and
high-definition video- is synchronized at the millisecond level and output at
200 hertz with a signal latency of just one millisecond.
Exploring the Practical Benefits of Using DAS Dex
The glove arrives factory-calibrated and
requires no additional setup. A user puts it on, and data collection begins
immediately. This matters because one of the biggest obstacles in training
robots has been the time and cost of collecting usable demonstration data.
Traditional teleoperation systems require the operator to control a robot
remotely while sensors record the robot's own movements. Those systems are
expensive, complex to set up, and often produce data that does not translate
cleanly to different robot hardware.
DAS Dex takes a different approach. Because it
captures the human hand directly, the data reflects natural human behavior
rather than a person struggling to operate unfamiliar robotic controls. The
glove stores data locally on 64 gigabytes of built-in storage and uploads it
seamlessly to the cloud when connected to Wi-Fi. The three-hour battery life
covers extended collection sessions. GenRobot AI's platform has accumulated
data across thousands of skills and scenarios spanning household, industrial,
logistics, and medical environments, and the company has committed to
open-sourcing portions of its data assets to support the broader robotics
research community.
How DAS Dex Fits Into GenRobot AI's Full System
The glove is one component of a larger product
line called GenDAS, which stands for data acquisition system. Other devices in
the family include DAS Ego, a head-mounted camera that captures first-person
perspective and spatial context; DAS Fingers, a fingertip-specific sensor
system; DAS Gripper, a handheld data collection device designed for simpler
grasping tasks; and DAS Controller, a synchronized control interface. Each
device captures a different aspect of human interaction with objects, and they
can be used independently or combined for richer, multi-perspective datasets.
Behind the hardware sits GenMatrix, an
AI-powered data governance platform that processes, organizes, and prepares the
collected data for use in training robotic AI models. The platform handles
compression, reducing raw data to roughly two percent of its original size, and
supports automatic cloud synchronization. GenRobot AI also operates GenADP, a
scalable data production pipeline designed for high-volume data collection
across multiple operators and locations simultaneously. The full stack-
hardware, software, and data services- is designed so that robotics companies
do not need to build their own data infrastructure from scratch.
Where Embodied AI Data Collection Fits in the Robotics Industry
The race to build capable humanoid and
industrial robots has created an equally urgent race to collect the data those
robots need to learn from. Every company developing a robotic hand, whether for
manufacturing, healthcare, logistics, or household applications, needs large
volumes of high-quality demonstration data showing how humans perform the tasks
the robot is expected to replicate. The quality, precision, and scale of that
data directly determine how well the robot performs.
Several approaches to collecting this data are
competing in the market. Some companies use vision-based systems that watch
human hands with cameras and reconstruct their movements computationally.
Others use teleoperation rigs where a human operator controls a robotic hand
remotely. DAS Dex represents a third approach: a wearable sensor glove that
captures human hand data directly, without cameras or robotic intermediaries.
Each method has trade-offs in cost, accuracy, ease of use, and how well the
collected data transfers to different robotic platforms. What all of them share
is a recognition that the quality of robotic manipulation will only be as good
as the human demonstration data used to train it, and that collecting that data
at scale remains one of the most significant practical challenges in bringing
capable robots to market.
You have exceeded your free limits for viewing our premium content
Please subscribe to have unlimited access to our innovations.
Get The Latest In Innovation
Join Our Newsletter
Where bold ideas find bold readers.
