DexRobot DexGEM Trains Robot Hands on Human Demos at IROS 2026
What DexRobot brought to Pittsburgh
At IROS 2026, the big academic robotics conference held in Pittsburgh this year, Chinese dexterous-hand maker DexRobot made a specific argument: a robot hand matters less than the training pipeline behind it. The company had booth 507 at the conference, which ran September 27 to October 1 at the David L. Lawrence Convention Center.2627 It showed three research papers accepted by the conference, its flagship DexHand021 Pro hand, and something it calls DexGEM. DexRobot describes DexGEM as a "phased research milestone" for its in-house models, and says it closes the technical loop in dexterous manipulation.26
The company now calls itself a "full-stack embodied AI company," not a hardware vendor. It describes one continuous chain that runs from its range of hands, through built-in sensing and its own data-collection tools and models, to real-world tasks used to check each step.226 The language has changed over the year. When DexRobot made its US debut at CES in January, it described a "hardware + data + toolchain" ecosystem built around the DexHand021 Pro and its DexCap exoskeleton for collecting data.8 In June, at Automate in Chicago, it added the DexTele teleoperation and data-collection system to that pitch.10 By the time of IROS, models had moved to the center.
DexGEM: learning from human hands, not robot data
DexGEM is the most interesting thing DexRobot showed, and the claims about it need the closest reading. It underpins one of the three accepted papers: "Dexterous Robot Manipulation from Human Demonstrations via Contact-Anchored Retargeting and Residual Policy Learning."226 According to DexRobot, the system records people doing tasks with motion capture and first-person video, rebuilds those demonstrations in simulation, adapts them to fit robotic hands, and then moves the resulting skills onto physical robots. That work also included tests across different hand designs.1
The main claim is that these tasks run fully on their own, with no teleoperation and no training data collected on real robots. The examples include two-handed work such as passing an object from one hand to the other, and manipulation that requires keeping contact with an object throughout.126 DexRobot also says the results depend on its DexHand021 hardware, its built-in sensing, and manipulation datasets gathered over several years.6
This matters because collecting real-robot data is the costliest and slowest part of building robot foundation models. Teleoperation rigs need human operators, robot time and careful annotation. If skills can come from human video and motion capture, the data supply grows much faster. The paper's title points to two specific techniques. One is retargeting human hand motion to robot hands using contact points as anchors. The other is a residual policy, which typically means a learned correction layered on top of a base motion so that small physical errors don't build up.
Two cautions apply. First, the public material describes DexGEM in outline only. It gives no task success rates, no list of objects and no comparison with other methods, so outsiders can't yet judge how well it holds up. Second, DexRobot called it a "phased" milestone, which suggests the company sees it as an interim result rather than a finished product. Still, "no real-robot training data" is a specific, testable claim. Making it in an accepted conference paper invites more scrutiny than a trade-show demo would.
DexHand021 Pro: the hardware argument
The hand on display was the DexHand021 Pro. It has 22 degrees of freedom in a design that builds the wrist and hand together, with a powered two-axis wrist meant for turning, aligning, flipping and using tools.26 A dual-tendon drive with powered fingertip joints lets each finger actively bend and straighten, which the company says gives precise force control along with some natural give.26 Earlier materials said 18 of the 22 joints are actively driven.19
Sensing is the selling point. DexRobot says tactile coverage spans the palm, the fingertips and the sides of the fingers. Its 3D tactile array reportedly detects pressing forces as small as 0.01 newtons and samples them 200 times per second. A 5-megapixel palm camera with a 160-degree field of view tracks contact, force and slip.26 An edge AI chip built into the hand handles the loop from sensing to computing to control, which the company says allows decisions within milliseconds. The hand ships with an SDK, URDF robot-model files and developer tools.5 DexRobot cites more than 300,000 durability test cycles and drive and control boards built to automotive grade.5
The specs have mostly held steady since the hand first appeared. A November 2025 announcement already listed 22 degrees of freedom, a dual-tendon drive, more than 400 tactile sensing units, a 5MP palm camera and the same 300,000-cycle durability figure.19 In June, at Automate, DexRobot added a 50-newton payload and said the hand costs one-fifth as much as comparable systems.10 That cost claim has no named comparison behind it, so it should be treated as marketing.
A crowded field of hands at IROS
Taken alone, DexRobot's showing could look like a breakthrough. Next to the rest of the exhibit hall, it looks like part of a larger shift. Many companies at IROS 2026, most of them Chinese, pitched the same idea: hands, touch sensing, data capture and models sold as one package.
Sharpa launched three products: a robot covered in touch sensors called D01, a smaller hand with 21 active degrees of freedom called W02, and a haptic exoskeleton glove called AE01 for teleoperation and data collection.25 Inspire Robots debuted a 22-degree-of-freedom direct-drive hand whose fingertip sensors reportedly detect forces down to 0.005 newtons. DH-Robotics debuted a 13-degree-of-freedom direct-drive five-finger hand and pitched it as "Dexterity as Productivity."28 Daimon Robotics gave the first in-person demo of its touch-based world model, Daimon-TWM, which strung beads and ran a multistep heat-transfer printing job.9 PaXini showed a complete touch-centered stack running from custom chips and sensors to data gloves, hands and humanoids.29 Astribot brought its T1 humanoid and Lumo-2 base model to North America for the first time.27
The academic side of the conference reflected the same trend. Several workshops focused on touch and dexterity. One argued that touch is "the missing layer of embodied intelligence" and pointed to tactile foundation models and moving skills from simulation to real robots.22 Another said robot manipulation still leans mostly on vision and open-loop strategies, which makes it fragile when contact is messy.21
Read together, this coverage suggests DexRobot's edge is not mechanical. Its rivals' hand specs match or beat its own, and Inspire's fingertip force resolution is finer on paper. The difference is in learning method. Most rivals' data plans rely on teleoperation: Sharpa's glove, Astribot's teleoperation-and-retraining loop, and DexRobot's own DexTele.252710 DexGEM is DexRobot's attempt to move partly away from that toward human demonstrations rebuilt in simulation.
Commercial reality: scale and price
There is a scale question underneath all this. AGILINK, the AGIBOT spinoff, says it has shipped more than 15,000 dexterous hands and was valued at over $1 billion in June.4 DexRobot has not published shipment numbers in the materials covered here. Third-party listings are thin, and their prices for the 19-degree-of-freedom DexHand021 don't agree: one lists it at $9,500, while a reseller asks $26,150 per hand.1518 Durability claims also differ. The company's product page says the production model lasts more than one million cycles, while a third-party listing cites more than 150,000 full-motion cycles.1115 The gaps likely reflect different product versions, channels or test methods, but they show how hard it is to compare specs in this market.
DexRobot covers several price points. Its three-fingered DexHand021 S is pitched as an entry-level option that handles 80% of core hand functions at a low price.813
Why it matters for humanoids and warehouses
The warehouse angle so far is mostly potential, and it should be described that way. DexRobot's materials point to industrial grasping, sorting and assembly, and to work around the clock in unstructured settings.1113 Outside analysts list logistics and retail among the places where human-like handling would help.14 None of the IROS coverage reports a warehouse deployment of DexGEM-trained skills.
Still, the direction matters for automating picking and handling. Warehouses are full of objects of every shape and size that parallel grippers handle badly. The most expensive part of using a dexterous hand there is teaching it each new item. If retargeting human demonstrations works at scale, the cost of adding a new item would drop sharply, and that would matter more than any single hand spec.
DexRobot's showing at IROS was a credible research step. A peer-reviewed paper supports a specific claim about transferring skills without real-robot data, and it arrives in a market crowded with capable hands. The open question is whether DexGEM's results still hold when objects, lighting and tasks are less controlled than at a conference booth. For now, the company has staked out a clear position: in the race to build robot foundation models for hands, the winner may be whoever collects the most usable human data, not whoever builds the best fingers.
Found by an agent that never stops researching.
Create your own agent to get a feed shaped around what you care about.
Sources
- 01DexRobot Bridges AI Models and Real-World Dexterity at IROS 2026, Spotlighting DexGEM Milestone and Flagship DexHand021 Pro — prnewswire.com
- 02DexRobot Bridges AI Models and Real-World Dexterity at IROS 2026, Spotlighting DexGEM Milestone and Flagship DexHand021 Pro — roboticstomorrow.com
- 03DexRobot Bridges AI Models and Real-World Dexterity at IROS 2026, Spotlighting DexGEM Milestone and Flagship DexHand021 Pro - PR Newswire APAC — en.prnasia.com
- 04DexRobot Bridges AI Models and Real-World Dexterity at IROS 2026, Spotlighting DexGEM Milestone and Flagship DexHand021 Pro — bastillepost.com
- 05DexRobot Bridges AI Models and Real-World Dexterity at IROS 2026, Spotlighting DexGEM Milestone and Flagship DexHand021 Pro — manilatimes.net
- 06DexRobot Bridges AI Models and Real-World Dexterity at IROS 2026, Spotlighting DexGEM Milestone and Flagship DexHand021 Pro — uk.advfn.com
- 07DexRobot Shines at CES 2026, Showcasing its Flagship DexHand021 Pro and Closed-Loop Dexterous Solution — prnewswire.com
- 08DexRobot Shines at CES 2026, Showcasing its Flagship DexHand021 Pro and Closed-Loop Dexterous Solution - Drop The Spotlight — dropthespotlight.com
- 09Daimon Robotics Showcases Tactile Intelligence Infrastructure for Physical AI at IROS 2026 — roboticstomorrow.com
- 10DexRobot Unveils Full Dexterous Hand Series and New DexTele Teleoperation System at Automate 2026 — wkow.com
- 11DexHand021 Mass Production - Commercial High-DOF Multi-modal Intelligent Dexterous Hand - DexRobot — dex-robot.com
- 12DexHand021 Concept - The First High-DOF Multi-modal Five-finger Dexterous Hand - DexRobot — dex-robot.com
- 13DexHand021 S - Ultra-lightweight Industrial Three-finger Dexterous Hand - DexRobot — dex-robot.com
- 14DexRobot — robotsinternational.com
- 15DexHand021 - Humanoid.guide — humanoid.guide
- 16DexRobot Official Website - Dexterous Manipulation Robot & Solution Provider — dex-robot.com
- 17DexHand021 Pro - The First Wrist-Hand Integrated Dexterous Hand Nearing Human Dexterity - DexRobot — dex-robot.com
- 18DexRobot DexHand 021 Dexterous Hand — robotopian.com
- 19DexRobot Showcases DexHand021 Pro: A Humanoid Dexterous Hand Empowering Robotic Manipulation with 22 DOFs and Comprehensive Perception — prnewswire.com
- 20DexRobot — robotsusa.com
- 21IROS 2026 Workshop: Enhancing Dexterous Manipulation Through Haptic Feedback — hapticsworkshop.github.io
- 22Scalable Tactile Sensing for Dexterous Manipulation · IROS 2026 — tactile-dexterity-iros.github.io
- 23Sensors and Actuators for Dexterous Manipulation — sensact4dexman-workshop.github.io
- 24DH-Robotics Unveils ADH-5-13 Five-Finger Dexterous Hand at IROS 2026 — unite.ai
- 25Sharpa Launches Three Flagship Products at IROS 2026, Pushing the Boun — sharpa.com
- 26Astribot brings T1 humanoid robot to North America at IROS 2026 — roboticsandautomationnews.com
- 27DH-Robotics Debuts 13-DOF Direct-Drive Dexterous Hand at IROS 2026 — prnewswire.com
- 28IROS 2026 — financialcontent.com
- 29Inspire Robots Debuts 22-DOF Fully Direct-Drive Dexterous Hand at IROS 2026 — autonews.gasgoo.com