AGILINK OmniHand 3 Ultra: 21 DoF, 630 g and a Robot Hand That Embroiders

AGILINK's OmniHand 3 Ultra is a direct-drive robot hand with 21 active degrees of freedom and a weight of 630 g, according to the company's own product page. On 30 September 2026 AGILINK released a video of the hand learning Suzhou silk embroidery from master craftswoman Fu Xianghong. No price or availability has been published, and every figure here is company-stated.

What is the AGILINK OmniHand 3 Ultra?

It is a human-sized, five-finger robot hand sold to robot builders and researchers, not to consumers. AGILINK lists it as the "Direct-driven Flagship Hand" of its OmniHand 3 line, next to the tendon-driven Ultra-T, the standard OmniHand 3, the OmniHand 3 Lite and the OmniPicker 3 gripper.

In plain terms: it is a hand with a small motor at every joint and a camera-style touch sensor in each fingertip, built for tasks where the grip has to change from moment to moment.

One naming wrinkle: the product page title says "Ultra-M" while the product is labelled "OmniHand Ultra". New Atlas notes that AGILINK has not said exactly which model appears in the embroidery video, but that it looks identical to the Ultra-M. This article treats the "Ultra" and the page's specs as one product.

What is AGILINK's thesis?

AGILINK bets that fingertip touch, not just finger count, is what separates a useful robot hand from a demo. The company describes itself as specialising in dexterous manipulation that combines hardware, tactile sensing and control for delicate and deformable objects.

In the release, CEO Qiao Tianjie says a single silk thread gives the fingertip dexterity "a visible measure". The company names cable handling, flexible-material assembly and precision-component manipulation as target applications. Embroidery is the showpiece, not the market.

How is AGILINK funded?

AGILINK says it was spun out of AGIBOT in January 2026. Its fourth funding round, in June, valued the company at more than US$1 billion, according to the release. The release names Hillhouse Ventures and Mirae Asset as investors and does not give the round size.

Qiao Tianjie is both AGILINK's CEO and an AGIBOT vice president, so the two companies remain closely linked. The company says it has shipped more than 15,000 dexterous hands. Its homepage claims 20,000+ products shipped, a gap that may reflect different dates. The release also says the OmniHand series won seven gold, four silver and three bronze medals across eight dexterous-hand events at the August 2026 World Humanoid Robot Games.

What are the OmniHand 3 Ultra's specifications?

The table lists what AGILINK's product page states as of 13 October 2026. None of it has been independently measured.

Two conflicts matter. The press coverage by New Atlas and Interesting Engineering gives a 3 kg steady hold and 8 kg lift, while the product page lists a 5 kg payload. And the page itself gives two repeatability figures. A buyer should ask for the datasheet.

How does the OmniHand 3 Ultra work?

Direct drive means each joint has its own actuator and encoder, with no tendons or intermediate transmission. AGILINK says joint states are read from each actuator's encoder. The trade-off is that actuators must fit inside the hand, which adds mass and bulk. Tendon designs such as AGILINK's own Ultra-T move actuators into the forearm to keep the hand light, per Humanoid.guide.

The sensing is the more unusual part. Each fingertip carries a vision-based tactile sensor, which New Atlas compares to "a camera peering at the inside of a soft pad". It reads contact from how the surface deforms. The company says this lets the hand measure force in three dimensions and detect deformation down to 0.08 mm, which would help it tell a slipping object from one it is merely brushing.

On the software side, the company says training combined teleoperation data with reinforcement learning, as reported by Interesting Engineering. The release does not mention the method, and no data volume has been published. The product page says the hand supports reinforcement-learning workflows through its simulation assets.

What did the embroidery demo actually show?

The video shows the hand separating silk thread, threading a needle, stretching fabric on a hoop and making stitches, taught by Fu Xianghong, a Suzhou embroidery inheritor. Per New Atlas, one thread is split into 16 finer strands, with one finger separating strands while the others hold tension.

Silk is a hard test because it bends, slips and changes shape with slight changes in contact. Twisting it needs opposing fingertips to stay in contact while the thread rotates. Interesting Engineering reports the final task was a twist-and-pick-up motion where the thread landed in a different place each time, so the hand had to adapt instead of replaying a fixed sequence.

That is a useful signal about the hand's touch sensing. It is not a benchmark. Interesting Engineering describes the demo as covering "basic embroidery techniques", and the footage is company-produced.

How does it compare with other robot hands?

The table sets the Ultra against two hands with published, sourced specs. All values are company-stated or relayed by the cited outlet.

The Sharpa Wave has one more active joint but is roughly twice as heavy, per CNX Software. AgiBot's OmniHand 2025 is a lighter, simpler hand with fewer active joints and a linkage mechanism. Both New Atlas and Interesting Engineering also compare the Ultra with LinkerBot's tendon-driven Linker Hand L30 Pro and Boston Dynamics' four-digit Atlas hand, which takes the opposite route to dexterity.

The practical difference: the Ultra pairs low weight with a high joint count and fingertip touch sensing, while the cheaper and simpler hands trade some of that for cost and size.

Where will people actually use this?

Not in a home, yet. The company's named targets are factory work. Likely uses, based on those targets:

• Cable handling: routing and seating flexible cables, which move unpredictably when pushed.
• Flexible-material assembly: fabric, film or soft parts that deform on contact.
• Precision-component manipulation: small parts where grip force must be exact.
• Research on robot learning: the Isaac and MuJoCo assets and Python SDK point to labs training hand policies.

A humanoid with this hand could eventually fold cloth, as in Dyna Robotics' Taku laundry robot, though AGILINK has made no such claim. For a sense of how fully built robots are priced, see the Unitree G1+. Not every dexterity route uses fingers on a hand, either: the Astribot T1 relies on cable-driven arms.

What's still unclear?

• Which model is in the video. AGILINK has not said whether the footage shows the Ultra or the Ultra-M, according to New Atlas.
• Payload and repeatability. The product page and press coverage disagree on payload (5 kg versus 3 kg hold and 8 kg lift), and the page gives two repeatability numbers.
• Training evidence. There is no published data volume, success rate, trial count or statement on video editing. The training method comes from a single outlet relaying the company.
• Generalisation. Silk handling does not prove the hand copes with other materials, and fine force control, New Atlas notes, likely will not be needed for most factory jobs.
• Shipment count. The release says 15,000+ hands, the homepage says 20,000+, and the homepage's "No.1" market-share claim carries no source or percentage.
• Cost and reliability. Price, availability and field durability beyond the 500K-cycle bench figure are unknown. For reference, Humanoid.guide lists the sibling OmniHand 3 Ultra-T at 12,000 USD as of 13 October 2026, a third-party listing and not an AGILINK price.

Should you buy it or wait?

If you build robots, request the datasheet and ask AGILINK to settle the payload and repeatability numbers before you design around the hand. The product page offers only a "Talk to Our Team" route. If you are a consumer, there is nothing to buy: this is a component for robot makers, and no price has been announced.

For anyone watching the industry, the embroidery clip is best read as a statement of intent about touch sensing. Independent measurements of grip, slip detection and repeatability would turn it into evidence.