Sharpa brings a touch-covered robot, a lighter 21-joint hand and a haptic glove to IROS, building its own data loop for dexterity

Sharpa's newest hand has one joint fewer than its last. The W02, unveiled at the IROS 2026 robotics conference in Pittsburgh, has 21 active degrees of freedom and weighs under 750 grams, the company said in a release dated 28 September, while its earlier SharpaWave hand has 22. The trade, a joint given up for weight and skin, says a lot about where dexterous robotics is heading.

Sharpa, founded in 2024 and headquartered in Singapore, launched three products at once: the D01 robot, the W02 hand and the AE01 haptic data glove. Read separately they are spec sheets. Read together they form a loop in which a human demonstrates a task through a glove, a robot repeats it with tactile hands, and the data trains Sharpa's models. The question this article answers is whether that loop is a moat or simply a bundle.

A robot wrapped in skin, built for human workspaces

The D01 combines Sharpa Wave hands, arms with seven degrees of freedom and electronic skin across the body, according to the release. Sharpa gives an arm payload-to-weight ratio of roughly 1:1, a maximum end-effector speed above 10.5 metres per second, a 1,000 Hz communication rate and end-effector repeatability of 0.2 millimetres. The skin samples at 100 Hz with a force resolution of 0.2 newtons, and human-scale spherical wrists are meant to let the machine work in spaces designed for people. The release does not say whether the D01 walks, rolls or sits on a fixed base, and gives no price or delivery date.

The W02 builds on Sharpa's W01 hand, the release says. It pairs Sharpa's Dynamic Tactile Array at the fingertips, rated from 5 millinewtons to 30 newtons at 1 millimetre spatial resolution, with electronic skin across fingers and palm covering 0.1 to 20 newtons at 5 millimetre resolution. It is rated IP54 and offers what Sharpa calls zero-gap grasping, closing fully enough to pick up chopsticks or fine cords.

The AE01 is a haptic exoskeleton glove for teleoperation and egocentric data capture. It has 22 degrees of freedom read by 22 encoders, 256 levels of vibrotactile feedback at each fingertip, position repeatability under 1 millimetre, and needs no calibration, according to Sharpa.

The company behind the hands

Sharpa says it was founded in 2024. A December 2025 release placed its global headquarters in Singapore, manufacturing and R&D in Shanghai and business operations in Mountain View, and said the SharpaWave hand, with 22 active degrees of freedom and more than 1,000 tactile pixels per fingertip, had entered mass production. At CES 2026 the company, which said it had more than 100 employees, showed a full-body robot called North and a hierarchical vision-tactile-language-action model called CraftNet, split into an interaction layer and a motion layer. In June 2026 NVIDIA chose dual Sharpa hands for its Isaac GR00T reference humanoid, a pairing ROBOTNESS reported at the time. Sharpa has not disclosed funding.

Sharpa has also quoted different fingertip sensitivities for the same Wave hand: 0.005 newtons in December 2025 and 0.02 newtons in June 2026. The company has not explained the difference, which may reflect different measurement methods.

How the W02 compares with rival hands

On maker-published figures, the W02 is light for its joint count. At a ceiling of 750 grams over 21 active joints it carries about 36 grams per actuated degree of freedom, by ROBOTNESS's calculation. Unitree's Dex5-1 weighs 1,100 grams with 16 active joints, about 69 grams each. Shadow Robot's Dexterous Hand has 20 actuated joints and a 4.3 kg full system, though that figure includes more than the hand. Wonik Robotics' Allegro Hand, a research staple, has 16 torque-controlled joints across four fingers.

Weight per joint matters because a hand sits at the end of an arm. Every gram there costs payload and dynamic performance, and Sharpa's claimed 1:1 payload-to-weight ratio on the D01 arm depends on light end effectors.

Why touch and a glove matter to robot learning

Most robot learning today relies on vision and joint positions. Contact, how hard a fingertip presses and where an object slips, is largely missing from training data because few hands can sense it and few teleoperation rigs can feed it back to the operator. The AE01 addresses the second half: an operator feels vibration at the fingertips while steering a robot, and the demonstration is recorded with matching hand poses. That is the input a model such as CraftNet needs.

There are open questions. Sharpa gave no prices, volumes or customers for the new products beyond saying they serve robotics companies, research institutions and food service. The D01's mobility and degrees of freedom outside the arms are not stated. The glove has 22 degrees of freedom and the new hand 21, so mapping between them requires retargeting, while the glove matches the 22-joint Wave one to one.