# Sharpa

> Source: https://aiwiki.ai/wiki/sharpa
> Updated: 2026-07-31
> Categories: Humanoid Robots, Robotics Companies
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Sharpa." aiwiki.ai, 31 Jul 2026. https://aiwiki.ai/wiki/sharpa
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Sharpa** is a Singapore-headquartered [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) [robotics](https://aiwiki.ai/wiki/robotics) company, founded at the end of 2024, that builds general-purpose robots and dexterous robotic hands with combined vision and touch sensing. Its flagship product is SharpaWave, a 22-degree-of-freedom human-scale robotic hand that won a CES 2026 Innovation Award; the company does not publish a price, and a figure of about USD 50,000 comes from reporting rather than from Sharpa.[1][2][5][7] Sharpa was established by Li Yifan (David Li), Xiang Shaoqing, and Sun Kai, the same three co-founders behind Hesai Technology, the Nasdaq-listed supplier of automotive [LiDAR](https://aiwiki.ai/wiki/lidar).[1][2][4] In May 2026, NVIDIA selected Sharpa's hands for the NVIDIA Isaac GR00T Reference Humanoid Robot, NVIDIA's first open humanoid robot reference design.[11] On July 30, 2026, [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind) published per-task results for [Gemini Robotics 2](https://aiwiki.ai/wiki/gemini_robotics_2) measured on SharpaWave hands mounted on an [Apptronik](https://aiwiki.ai/wiki/apptronik) [Apollo 2](https://aiwiki.ai/wiki/apollo_2), the most detailed third-party evaluation the hand has received.[16]

| Sharpa | |
| --- | --- |
| General information | |
| **Full name** | Sharpa Pte. Ltd. |
| **Founded** | 2024 (end of 2024) |
| **Founders** | Li Yifan (David Li), Xiang Shaoqing, Sun Kai |
| **Headquarters** | Singapore |
| **Other offices** | Shanghai (R&D and manufacturing); Mountain View, California (business operations) |
| **Industry** | [Robotics](https://aiwiki.ai/wiki/robotics), [Artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) |
| **Employees** | 100+ (as of early 2026) |
| **Products** | Sharpa Wave (SharpaWave) dexterous hand, [North](https://aiwiki.ai/wiki/sharpa_north) humanoid robot, CraftNet AI model |
| **Key technology** | Dynamic Tactile Array (DTA), visuo-tactile sensing, dexterous manipulation |
| **Notable** | Hands selected for the [NVIDIA Isaac GR00T](https://aiwiki.ai/wiki/nvidia_isaac_gr00t) Reference Humanoid Robot (2026); evaluation platform for Google DeepMind's Gemini Robotics 2 multi-finger dexterity results (2026) |
| **Website** | [sharpa.com](https://www.sharpa.com/) |

Sharpa specializes in human-scale dexterous manipulation, [tactile sensing](https://aiwiki.ai/wiki/tactile_sensing), and autonomous fine-manipulation systems. The company is best known for SharpaWave, which won a CES 2026 Innovation Award, and for North, a humanoid robot that played fully autonomous ping-pong against human opponents at the same trade show.[1][2][3] Sharpa operates as an independent venture with no equity or business control relationship with Hesai. Its global headquarters and corporate functions sit in Singapore, while its manufacturing and most of its research and development activities are concentrated in Shanghai. A separate office in Mountain View, California, supports business operations and customer engagement in North America.[2][3][4]

Since mid-2026 the company has branded the hand as **Sharpa Wave** on its own product pages, while the older single-word spelling **SharpaWave** persists in its 2025 press releases, in NVIDIA's materials, and in Google DeepMind's figure captions. Both names refer to the same product, and this article uses them interchangeably.[12][16][11]

## What is Sharpa?

Sharpa (legally Sharpa Pte. Ltd.) is an [AI robotics](https://aiwiki.ai/wiki/ai_robotics) startup that develops high-performance general-purpose robots and core robotic components, with a focus on contact-rich "fine" manipulation: the regime in which a task succeeds or fails on sub-centimeter accuracy and precise control of grip force. The company's central technical bet is visuo-tactile sensing, an approach Sharpa describes as "feel by seeing," in which a miniature camera and a dense pressure-sensor array in each fingertip are fused so a robot hand can both see and feel what it is touching.[2][5][8] Sharpa's stated mission is to "manufacture time by making robots useful," framing its products as tools that take over repetitive or strenuous work.[2][8]

The company frames its own target in terms of the residual failure rate rather than the headline success rate. Its overview page argues that "a task completed at 99% is still unfinished" and that the useful work sits in the last one percent, giving tying off a trash bag and clipping a wire into its slot on an assembly line as examples.[28] Founder David Li made the same point to Singapore's Economic Development Board in June 2026: "We are not satisfied with lab demonstrations that succeed 90 per cent of the time."[31]

## Who founded Sharpa?

Sharpa was incorporated in Singapore at the end of 2024 by three founders who had previously co-founded Hesai Technology, the Nasdaq-listed [LiDAR](https://aiwiki.ai/wiki/lidar) manufacturer based in Shanghai. The trio consists of Li Yifan, who serves as chief executive of Hesai; Xiang Shaoqing, Hesai's chief technology officer; and Sun Kai, the company's chief scientist. According to reporting by Chinese technology outlet 36Kr, the founders treat Sharpa as a "second entrepreneurial project" and serve in strategic advisory and direction-setting capacities at Sharpa rather than holding day-to-day operational positions there.[2][4]

That description needs one update. From mid-2026 onward, Sharpa's own materials and Singapore government press releases identify David Li by the title Founder of Sharpa: NVIDIA's May 2026 reference-robot announcement, JTC Corporation's April 28, 2026 memorandum-of-understanding release, and a June 30, 2026 Singapore Economic Development Board interview all quote him under that title.[11][32][31] Day-to-day operations remain with an externally recruited team, and Sharpa has not published an executive roster.

The founders have publicly described their interest in [robotics](https://aiwiki.ai/wiki/robotics) as a natural extension of Hesai's expertise in spatial perception. Hesai's core competence is in [LiDAR](https://aiwiki.ai/wiki/lidar) sensors that allow autonomous machines to perceive three-dimensional environments, and Sharpa's leadership has suggested that the long-term market for robotics-grade LiDAR may eventually exceed the market for automotive LiDAR. Despite this strategic alignment, Sharpa is structured as a fully separate corporate entity, recruits most of its core team externally rather than transferring staff from Hesai, and operates with its own brand, balance sheet, and product roadmap.[2][4]

Chinese technology outlet Leiphone reported in March 2026 that Hesai's board did not initially support expanding away from LiDAR, so the robotics effort was incorporated as a separate entity outside the listed company. The same report noted the governance objection this creates: profits from Sharpa would not flow to Hesai shareholders who did not participate in it, and the three founders would be splitting attention across two companies. Leiphone contrasted the arrangement with rival LiDAR maker RoboSense, which kept its robotics business inside the listed entity.[17] Sharpa has not commented publicly on that criticism.

### What are the founders' backgrounds?

Li Yifan, who is also referred to publicly as David Li, holds a bachelor's degree in precision instruments and mechanology from [Tsinghua University](https://aiwiki.ai/wiki/tsinghua_university) and a Ph.D. from the University of Illinois Urbana-Champaign, where his doctoral research focused on humanoid robotics and high-precision motion control. Before co-founding Hesai, he worked as a principal engineer at Western Digital Corporation in Silicon Valley, where he specialized in high-speed, high-precision disk-head motion control. Xiang Shaoqing earned his undergraduate degree at Tsinghua in the same precision instruments and mechanology program and then completed dual master's degrees at [Stanford University](https://aiwiki.ai/wiki/stanford_university) in electrical and mechanical engineering, after which he worked on iPhone system integration at Apple. Sun Kai studied mechanical engineering at Shanghai Jiao Tong University and earned a Ph.D. in mechanical engineering from Stanford University. The three founders' overlapping expertise in mechatronics, motion control, and product engineering forms the technical backbone of Sharpa's hardware program.[2]

36Kr adds one detail on Sun Kai: his Shanghai Jiao Tong undergraduate degree was in the School of Mechanical Engineering, majoring in thermal energy and power engineering, before the Stanford mechanical engineering doctorate.[2]

## When did Sharpa launch its products?

Sharpa publicly introduced its first product, the SharpaWave [dexterous hand](https://aiwiki.ai/wiki/dexterous_hand), in May 2025. Over the following months the company demonstrated the hand at industry events and shipped early units to research customers. In October 2025, Sharpa held a demonstration at IROS 2025, the IEEE/RSJ International Conference on Intelligent Robots and Systems, held in Hangzhou, China. At IROS 2025 the SharpaWave performed tactile sensing tasks, manipulated playing cards, took autonomous photographs, and dealt and played blackjack with attendees.[5][6]

The May 2025 debut was at ICRA 2025, the IEEE International Conference on Robotics and Automation, where the hand drew attention for peeling an eggshell and using scissors. 36Kr, which covered that appearance, described the company as having stayed deliberately unidentified: investors who tried to reach it were turned down, and one told the outlet that as of early 2026 he still did not know who ran it.[2]

On October 16, 2025, Sharpa announced that it had begun shipping the SharpaWave to customers. On December 16, 2025, the company announced that the product had entered mass production, accompanied by automated reliability and endurance testing systems designed to validate thousands of microscale gears, motors, and sensors per unit. On November 6, 2025, the SharpaWave was named a CES 2026 Innovation Awards Honoree in the Robotics category, drawn from a field of more than 3,600 submissions to that year's program. The award listing describes the product as a hand for robotics companies, academic institutions, and research labs, with 22 active degrees of freedom at 1:1 human size and scale.[6][7][8][38]

| Date | Milestone |
|---|---|
| End of 2024 | Sharpa incorporated in Singapore |
| May 2025 | SharpaWave dexterous hand publicly introduced |
| October 2025 | Live demonstration at IROS 2025, Hangzhou |
| October 16, 2025 | SharpaWave begins shipping to customers |
| November 6, 2025 | SharpaWave named a CES 2026 Innovation Awards Honoree (Robotics) |
| December 16, 2025 | SharpaWave enters mass production |
| January 6-9, 2026 | North humanoid and CraftNet model debut at CES 2026 |
| February 2026 | Sharpa hands appear on CCTV's Spring Festival Gala, rolling walnuts[17] |
| February 25, 2026 | Tacmap tactile-simulation paper posted to arXiv with NVIDIA and HKUST co-authors[23] |
| April 28, 2026 | Memorandum of understanding with Singapore's JTC Corporation for Punggol Digital District[32] |
| May 31, 2026 | Sharpa hands chosen for the NVIDIA Isaac GR00T Reference Humanoid Robot |
| June 2026 | Four independent research groups publish work using Sharpa Wave hands[19][20][21][22] |
| July 30, 2026 | Google DeepMind publishes Gemini Robotics 2 multi-finger results measured on SharpaWave hands[16] |
| August 2026 (announced) | North scheduled to begin working shifts at a Dairy Queen store in Shanghai[30] |

### What did Sharpa show at CES 2026?

Sharpa achieved broad public visibility at CES 2026 in Las Vegas, held January 6 through 9, 2026, where it exhibited at booth 9251 in the LVCC North Hall. In addition to displaying the SharpaWave hand, the company introduced its first full-body humanoid robot, [North](https://aiwiki.ai/wiki/sharpa_north), and a new artificial intelligence model called CraftNet. The North robot played fully autonomous matches of ping-pong against human opponents during eight-hour daily live sessions across the four days of the show. The booth's other demonstrations included [photography](https://aiwiki.ai/wiki/photography) (more than 2,000 instant photos captured during the show), assembly of paper windmills (more than 300 windmills assembled, with chains of more than 30 consecutive successful steps), and card dealing using multimodal vision and language processing. The demonstrations generated significant social media engagement and were widely covered by trade press.[8][9][10]

According to Bowei Liu, a Sharpa representative who spoke with reporters at the event, the North system reacts to changes in its environment in approximately 0.02 seconds. Alicia Veneziani, who serves as Sharpa's global vice president of go-to-market and president of Europe, told reporters that while many humanoid robots can already demonstrate athletic feats such as dancing or backflips, manipulation of objects with human-like precision remains the harder unsolved problem the company is targeting.[9][10]

## What does Sharpa make?

Sharpa's product line spans a dexterous hand (SharpaWave), a full-body humanoid robot (North), and a control model (CraftNet) that ties them together.

### SharpaWave dexterous hand

The SharpaWave is Sharpa's flagship product and the foundation of its product portfolio. It is a human-scale, biomimetic robotic hand designed to replicate the size, range of motion, and tactile sensitivity of an adult human hand. The product entered mass production in December 2025 and won a CES 2026 Innovation Award.[5][6][7]

The table below reproduces Sharpa's own published figures. Where the company's current user manual and its earlier marketing material disagree, the manual is used and the earlier figure is noted in the following paragraphs.

| Specification | Value | Source |
|---|---|---|
| Active degrees of freedom | 22 | Sharpa user manual[13] |
| Form factor | 1:1 human hand scale; palm-width to hand-length ratio about 0.618 | Sharpa product page[12] |
| Weight | 1.3 kg | Sharpa user manual[13] |
| Dimensions | 208 mm x 90 mm x 50 mm (height x width x thickness) | Sharpa user manual[13] |
| Mounting interface | Flange, wrist, or internal | Sharpa user manual[13] |
| Maximum active fingertip force | 20 N | Sharpa user manual[13] |
| Minimum grasp diameter | 10 mm | Sharpa user manual[13] |
| Fingertip position repeatability | Plus or minus 1 mm | Sharpa user manual[13] |
| Control frequency | 500 Hz | Sharpa user manual[13] |
| Operating speed | More than 4 Hz across all gestures | Sharpa product page[12] |
| Payload | 40 kg (term not defined by Sharpa) | Sharpa product page[12] |
| Tactile pixels per fingertip | More than 1,000 | Sharpa product page[12] |
| Tactile force resolution | 20 mN | Sharpa user manual[13] |
| Tactile measurement range / maximum load | 30 N / 50 N | Sharpa user manual[13] |
| Tactile spatial resolution | 1 mm | Sharpa user manual[13] |
| Tactile outputs | 6-DoF force, deformation map, contact point, raw image | Sharpa user manual[13] |
| Tactile sampling rate | 30 Hz standard mode; up to 180 Hz high-performance mode | Sharpa user manual[13] |
| Tactile latency | 50 ms standard mode; 20 ms high-performance mode | Sharpa user manual[13] |
| Torque sensors | 5, one per finger (thumb CMC, others MCP); 5 N-m full scale on the thumb, 3 N-m elsewhere; 500 Hz | Sharpa user manual[13] |
| Communication | Ethernet (100BASE-TX / 1000BASE-T) | Sharpa user manual[13] |
| Power | 18 to 28 V DC; 15 W static, 100 W average, 180 W peak (10 ms transient) | Sharpa user manual[13] |
| Environmental | 0 to 45 degrees C operating; IP20; humidity below 60% RH | Sharpa user manual[13] |
| Nominal lifetime | More than 1,000,000 cycles unloaded; tactile sensors more than 100,000 cycles at 40 N | Sharpa user manual[13] |
| Sensing technology | Dynamic Tactile Array (DTA) with miniature cameras | Sharpa product page[12] |
| Indicative price | Not published by Sharpa; about USD 50,000 reported by 36Kr in January 2026 | 36Kr[2] |

Several figures in circulation, including in earlier versions of this article, come from Sharpa's November 2025 CES award press release rather than from the datasheet, and the two disagree. The press release claims "0.005 N precision" for force sensing; the current manual specifies a tactile force resolution of 20 mN, that is 0.02 N, and the product page gives the same 0.02 N figure. The 0.005 figure does appear in the manual, but as the root-mean-square noise floor of the thumb's *torque* sensor in newton-metres, not as a force precision in newtons.[5][12][13] Similarly, "180 frames per second" is the high-performance raw-image mode; the default standard mode runs at 30 Hz. The hand's mass is 1.3 kg rather than the approximately 1,200 g previously reported, although the link masses in Sharpa's own published URDF simulation model sum to about 1.25 kg.[13][15] Maximum fingertip force is 20 N, not "more than 20 N."

**How the 22 degrees of freedom decompose.** Degree-of-freedom counts for robot hands are frequently inflated by counting passively coupled joints that share one motor, so the decomposition matters more than the total. Sharpa's manual and its published URDF and MJCF models give the full breakdown: the thumb has five joints (carpometacarpal flexion-extension and abduction-adduction, metacarpophalangeal flexion-extension and abduction-adduction, and interphalangeal), the index, middle, and ring fingers have four each (metacarpophalangeal flexion-extension and abduction-adduction, proximal interphalangeal, and distal interphalangeal), and the little finger has five, adding a carpometacarpal joint. That is 5 + 4 + 4 + 4 + 5 = 22.[13][15] The published MuJoCo model declares 22 independent position actuators, one per joint, with no tendon or equality constraints and no mimic joints in the URDF, and Sharpa's software development kit exposes firmware version queries over exactly 22 individually addressed motors.[13][15]

An independent group confirmed the same thing from the outside. The UC Berkeley-led T-Rex paper, which ran its real-world experiments on two Sharpa Wave hands, reports that it had to modify a baseline policy written for 21-degree-of-freedom hands with "several mechanically coupled joints masked out during prediction," and instead predicted all finger joints directly, which the authors say was "enabled by the fully actuated hardware design."[19] Sharpa's hand therefore has 22 actuated degrees of freedom and 22 total joints, with no underactuation, which is unusual at this size.

Leiphone described the SharpaWave as the first industrial-grade dexterous hand to reach 22 degrees of freedom with direct motor drive rather than cable transmission, and quoted Mu Yao, a young researcher at Shanghai AI Laboratory, saying that fitting visuo-tactile sensing into fingertips of roughly human size requires direct drive with in-house motors small enough and strong enough to fit.[17] Sharpa's own manual does not describe the transmission, so "direct drive" should be read as a characterization by informed third parties rather than a published specification. Sharpa has not published motor type, gear ratios, backdrivability figures, or a bill of materials, and it does not publish a price.

Each fingertip integrates a miniature camera alongside more than 1,000 tactile pixels, giving the system both visual and pressure-based feedback at the point of contact. The hand provides 6-dimensional force detection at each fingertip alongside a deformation map, a contact point, and the raw sensor image, supporting dynamic grip control and slip prevention during manipulation.[1][3][5][6][13]

From a maintainability perspective, the SharpaWave is designed in a modular configuration in which individual fingers can be replaced without rebuilding the entire hand. Sharpa positions this as an advantage for industrial customers compared with unibody alternatives, since damaged fingers can be swapped out rather than triggering a full unit replacement. Pricing was reported by 36Kr to be in the tens of thousands of US dollars, with a figure of approximately USD 50,000 cited by industry coverage.[2][6] That figure traces to a single January 2026 36Kr report attributing it to embodied-intelligence practitioners who wanted to buy one, and to the same report's observation that units were hard to obtain: one model company was told at the end of 2025 that there was no stock and that it could only run experiments at Sharpa's Shanghai office.[2] Sharpa's product page carries no price and directs buyers to a sales form.[12]

Demonstrated capabilities of the SharpaWave include cracking and peeling eggs, cutting with scissors, manipulating cards, folding paper, operating industrial tools, and playing table tennis when mounted on a robotic platform. The hand is sold to global technology companies and top research universities as both a standalone manipulator and as a building block of the larger North humanoid system.[2][5][6][7]

### North humanoid robot

[North](https://aiwiki.ai/wiki/sharpa_north) is Sharpa's first full-body autonomous [humanoid robot](https://aiwiki.ai/wiki/humanoid_robot), unveiled at CES 2026. North is built around a pair of SharpaWave dexterous hands attached to a torso with extensive neck, shoulder, and waist range of motion, and it moves on a wheeled base that can position the upper body to reach and react to dynamic targets. The system is intended as a full-stack manipulation platform capable of learning multiple fine-manipulation tasks rather than being purpose-built for a single application.[8][9]

Sharpa's product page gives North a total of 67 degrees of freedom and describes it as built to run shifts autonomously in service and industrial settings. Sharpa does not publish the breakdown, although two SharpaWave hands would account for 44 of that total.[27]

Notable demonstrations at CES 2026 include autonomous ping-pong rallies against human opponents, with reaction times reported at approximately 0.02 seconds, autonomous photography with positioning accuracy described in the millimeter range, autonomous assembly of paper windmills with sequences of more than 30 consecutive successful steps, and autonomous card dealing. Sharpa has indicated that a production version of North is expected in mid-2026 and is targeted at retail, hospitality, food-service, and domestic environments.[8][9][10]

That mid-2026 target slipped. In the June 30, 2026 Singapore Economic Development Board interview, David Li said North "will become available to enterprises by the end of 2026" and would begin executing real-world shifts starting with the service industry.[31] Sharpa opened an enquiry form on the North product page rather than a price list or order page, and as of July 31, 2026 it has published no North specification sheet, price, runtime, payload, or reach figure.[27]

The first announced commercial deployment is a Dairy Queen store in Shanghai. Sharpa says North will complete a full shift behind the counter of an operating store using the existing equipment, ingredients, and process unmodified, with each order running as a sequence of 50 to 60 consecutive manipulation steps performed end to end. The tasks it names are holding a metal ring against a paper cup with thumb and index finger throughout blending, opening a cabinet door with a narrow human-designed handle, tracking its own progress through three identical-looking scoops of Oreo crumbs, and performing the chain's signature upside-down flip of the finished Blizzard without crushing the cup or dropping the product. Sharpa says the flip in particular requires tactile-driven control that no teleoperation device can currently achieve, so North must run autonomously to perform it. The company gives August 2026 as the launch month.[30]

### CraftNet AI model

Alongside North, Sharpa introduced CraftNet at CES 2026, an end-to-end hierarchical vision-tactile-language-action (VTLA) model that controls the company's manipulation platforms. CraftNet is structured as a two-layer control system. The lower layer, which Sharpa calls the "Interaction Brain," handles fine-grained contact responses such as adjusting grip force in reaction to slip or surface compliance. The upper layer, the "Motion Brain," handles coordinated whole-body motion, planning, and task sequencing. The architecture is positioned by the company as targeting what it calls "last-millimeter" precision, the regime in which contact-rich manipulation tasks are decided by sub-centimeter accuracy. Sharpa has indicated that CraftNet will be released as a series of phased updates rather than as a single discrete release.[8]

Sharpa's current CraftNet page adds the operating frequencies and a third layer above the two. A System 2 "Reasoning Brain," a vision-language model running at roughly 1 Hz, passes semantic intent down to the System 1 Motion Brain at roughly 10 Hz, which passes coarse actions and physical intent to the System 0 Interaction Brain at roughly 100 Hz, which emits the fine actions. Sharpa defines CraftNet itself as the combination of System 0 and System 1, with System 2 sitting above it as the reasoning layer.[24] The naming follows the [vision-language-action model](https://aiwiki.ai/wiki/vision_language_action_model) convention of separating a slow deliberative layer from a fast reactive one, with the tactile loop as the fastest tier.

## How does Sharpa's hand technology work?

### Dynamic Tactile Array (DTA)

Sharpa's principal technological differentiator is its **Dynamic Tactile Array** (DTA) sensing system. DTA combines vision-based sensing in the form of a miniature camera embedded in each fingertip with a dense array of more than 1,000 tactile pixels per fingertip. A neural-network-based algorithm fuses the data from these two modalities to produce force estimates with 0.005 N precision across a 0 to 30 N range, sampled at 180 frames per second with sub-millimeter spatial resolution. The system is therefore able to interpret contact information across a wide dynamic range, supporting tasks that span delicate operations such as cracking an egg through to firm grasping of heavier industrial tools.[5][6]

Sharpa's current datasheet restates those numbers differently, and the datasheet should be preferred: force resolution 20 mN over a 30 N measurement range with a 50 N maximum survivable load, 1 mm spatial resolution, and 30 Hz sampling by default with a 180 Hz high-performance mode that streams only the raw image and requires gigabit Ethernet. Latency is 50 ms in standard mode and 20 ms in high-performance mode.[13] In academic papers co-authored with Sharpa, the sensor is referred to as Sharpa DTC and classified as an image-type (that is, camera-based) tactile sensor, alongside GelSight-Mini and similar devices, rather than as an electrode array.[20]

Sharpa frames this approach as "feel by seeing," or visuo-tactile sensing, in which the optical and mechanical channels reinforce each other rather than acting as independent inputs. In practice, this allows the SharpaWave to detect incipient slip events and adjust grip in real time, and it allows downstream policies such as CraftNet to reason jointly about visual and tactile evidence at the point of contact.[5][8]

The simulation counterpart is Tacmap, a tactile-simulation framework published in February 2026 by authors from Sharpa, the Hong Kong University of Science and Technology, and NVIDIA. Tacmap represents contact as a volumetric penetration-depth map in simulation and learns a mapping from raw tactile images to the same depth-map representation in the real world, so that both domains share one geometric language. The authors argue this avoids the two standard failure modes of tactile simulation, namely simplified geometric projections that are physically unrealistic and finite-element methods that are too slow for large-scale [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning). They validate it with an in-hand rotation policy trained only in simulation and transferred zero-shot to physical hardware.[23] Sharpa publishes Tacmap as an [Isaac Lab](https://aiwiki.ai/wiki/isaac_lab) plugin.[29]

### Software stack and developer tools

Sharpa supplies the SharpaWave with an open, developer-oriented software stack. The hand is shipped with SharpaPilot, a control application that includes example workflows for [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning), and the company maintains a public documentation portal under the sharpa-robotics GitHub organization.[14] The platform is compatible with widely used robotic simulation environments including [NVIDIA Isaac Lab](https://aiwiki.ai/wiki/isaac_lab), Isaac Gym, PyBullet, and [MuJoCo](https://aiwiki.ai/wiki/mujoco), allowing customers to develop and benchmark policies in simulation before deploying to physical hardware. The SharpaWave communicates with host computers using a standard Ethernet interface, simplifying integration into existing research and industrial setups.[5][6][8]

The current published stack is narrower and better documented than that list suggests. The control application is now written as two words, Sharpa Pilot, and the compatibility line on the product page names [ROS](https://aiwiki.ai/wiki/ros), Isaac Sim, and MuJoCo.[12] Sharpa's downloads page links ten public repositories: Sharpa Pilot, Sharpa Wave firmware, the Sharpa Wave SDK, a Manus MetaGloves Pro tracking and retargeting SDK, URDF/USD/MJCF assets, tactile sensor assets, the Tacmap Isaac Lab plugin, an OpenCAD mechanical library covering the mounting adapter and simplified models, a reinforcement-learning example in Isaac Lab, and a community forum.[29] The SDK ships C++ headers and a Python extension for Python 3.10 through 3.12 on x86-64 and ARM64 Linux, with samples for C++, Python, and ROS.[39] The asset repository provides left, right, and dual-hand models in URDF, MJCF, and USD, each with wrist, flange, and floating-base variants, plus a separate USD tuned to the hardware's MIT control mode.[15]

### Manufacturing and reliability engineering

To support its move into mass production, Sharpa developed automated reliability and endurance testing systems that validate the performance of the thousands of microscale gears, motors, and sensors that make up each SharpaWave unit. Sharpa describes these testing systems as "highly automated" and as a key prerequisite for delivering consistent units at scale. The hand's modular design, in which individual fingers can be replaced rather than the whole unit, is positioned as a serviceability advantage that lowers downtime and total cost of ownership for industrial customers compared with unibody alternatives.[6][7]

Sharpa now publishes six specific durability results, all of them its own internal test claims rather than third-party certifications:

| Test | Published result |
|---|---|
| Press cycles | Validated through 2,500,000 cycles[12] |
| Fingertip friction | Verified over 4,000 m of travel distance[12] |
| Impact | Hundreds of cycles without damage[12] |
| Mechanical shock | 3,200 impact cycles at 30 g acceleration[12] |
| Extreme temperature cycling | More than 1,000 hours of continuous operation[12] |
| Automatic protective clench | Responds in 0.10 s[12] |
| Nominal joint lifetime | More than 1,000,000 cycles unloaded; tactile sensors more than 100,000 cycles at 40 N[13] |

The friction figure is worth flagging because it is widely misquoted: Sharpa's own page and specification sheet both give 4,000 m, while at least one trade write-up rendered it as 4,000 km.[12][18] Durability claims of this kind matter more for dexterous hands than for most robot subsystems, because tendon-driven designs historically wear out fast, and cycle-life numbers have become the standard way suppliers argue that their transmission choice survives production duty.

## Google DeepMind's Gemini Robotics 2 evaluation

Main article: [Gemini Robotics 2](https://aiwiki.ai/wiki/gemini_robotics_2)

On July 30, 2026, Google DeepMind announced Gemini Robotics 2, the second numbered generation of its robotics [foundation model](https://aiwiki.ai/wiki/foundation_model) family, and published per-task success rates for it. The multi-finger dexterity results were measured on an Apptronik [Apollo 2](https://aiwiki.ai/wiki/apollo_2) humanoid fitted with SharpaWave hands, which Google DeepMind describes as "the five-fingered, 22 degree-of-freedom SharpaWave hand."[16] This is the most detailed independent measurement published on the hand to date.

Two points frame everything that follows. First, these numbers measure Google DeepMind's model, not the hand: they say how often a particular learned policy completed a particular household task on this hardware, and a different policy on the same hardware would produce different numbers. Second, Google DeepMind published no comparison against any other hand for these tasks, so nothing here establishes that SharpaWave is better or worse than an alternative.

| Task | Success rate | Platform |
|---|---|---|
| Unscrew bulb | 92% | Apollo 2 with SharpaWave hands |
| Tie trash bag | 44% | Apollo 2 with SharpaWave hands |
| Ziplock | 40% | Apollo 2 with SharpaWave hands |
| Screw bulb | 36% | Apollo 2 with SharpaWave hands |
| Dustpan | 32% | Apollo 2 with SharpaWave hands |

Google DeepMind reports these five as individual task results rather than category averages, unlike the whole-body and gripper bars published in the same figure. Its own caption states that while the model reaches medium to high success on whole-body and gripper-based dexterous tasks, multi-finger dexterous manipulation remains challenging.[16]

The single most informative number in the set is the 56-point gap between unscrewing a bulb (92%) and screwing one in (36%). Both actions use the same hand on the same fitting. Removing a bulb means finding a purchase that already exists and rotating; installing one means aligning a thread that is not yet engaged and seating it by feel, which is exactly the class of task the hand's tactile stack is built for and exactly where it still fails two times in three. The dustpan result (32%) is the lowest in the set.

A second result matters more for the hand's positioning than any individual percentage. Google DeepMind states that one model checkpoint drove three different embodiments: Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and a Franka Duo with a Robotiq two-fingered parallel gripper.[16] The same weights therefore controlled a 22-degree-of-freedom five-fingered hand and a standard two-fingered parallel gripper. For Sharpa this cuts both ways: it demonstrates that a frontier [robot foundation model](https://aiwiki.ai/wiki/robot_foundation_model) can address the hand's full joint space without hand-specific training, and it also shows that the model's highest published success rates in the same release came from the gripper, at 89.6% on precise insertion.

One inconsistency in Google DeepMind's own material is worth noting for anyone matching the charts to the text: the multi-finger chart's subtitle reads "Apollo with Sharpa hands" while the figure caption below it reads "SharpaWave hands." Both refer to the same hardware.[16]

Google DeepMind also used Apollo 2 for the safety evaluations published with the release, and the Inspire hands rather than the SharpaWave for its general whole-body picking results, so the division of labour in the release was gross manipulation on the lower-count hand and fine motor work on the higher-count one.[16]

## Independent research use and third-party adoption

Outside Google DeepMind, the Sharpa Wave has become a recurring hardware choice in published manipulation research, mostly during the first half of 2026. The following uses are each documented in a dated paper or a named customer account.

| User | What it was used for | Date | Source |
|---|---|---|---|
| [NVIDIA](https://aiwiki.ai/wiki/nvidia), [Unitree](https://aiwiki.ai/wiki/unitree) | Dual Sharpa Wave hands specified as the end effectors of the NVIDIA Isaac GR00T Reference Humanoid Robot | May 31, 2026 | NVIDIA newsroom[11] |
| Sharpa, HKUST, NVIDIA | Tacmap tactile simulation framework, validated by zero-shot [sim-to-real](https://aiwiki.ai/wiki/sim_to_real) transfer of an in-hand rotation policy | February 25, 2026 | arXiv 2602.21625[23] |
| [UC Berkeley](https://aiwiki.ai/wiki/uc_berkeley) and collaborators (T-Rex) | Two 22-DoF Sharpa Wave hands on a [Dexmate](https://aiwiki.ai/wiki/dexmate) Vega-1 bimanual robot; 100-hour tactile dataset; 12 contact-rich tasks | June 15, 2026 | arXiv 2606.17055[19] |
| [Tsinghua University](https://aiwiki.ai/wiki/tsinghua_university), Shanghai Qi Zhi, Sharpa, SJTU, UC Berkeley, [ETH Zurich](https://aiwiki.ai/wiki/eth_zurich) and others (FTP-1) | Sharpa North and a Sharpa-plus-Dexmate setup as two of the evaluation embodiments for a generalist tactile policy; Sharpa contributed a 4,000-demonstration dataset | June 11, 2026 | arXiv 2606.13102[20] |
| UC Berkeley (Do as I Do) | 22-DoF Sharpa Wave hands on Universal Robots UR3e arms, retargeting monocular human video to a dexterous hand | June 17, 2026 | arXiv 2606.19333[21] |
| Nanyang Technological University, Singapore (DexTeleop-0) | Two Sharpa Wave hands on twin [Universal Robots](https://aiwiki.ai/wiki/universal_robots) UR7e arms, 56 degrees of freedom in total, for force-aware bimanual [teleoperation](https://aiwiki.ai/wiki/teleoperation) | June 22, 2026 | arXiv 2606.23431[22] |
| NVIDIA (EgoScale) | 22-DoF Sharpa hand as the target embodiment for policies pretrained on about 20,854 hours of egocentric human video | 2026 | Sharpa research blog[25] |
| Cornell University and Stanford University (SimToolReal) | Sharpa Wave as the deployment hardware for a zero-shot object-centric tool-use policy | 2026 | Sharpa research blog[26] |

Three of these deserve more detail because they say something about the hardware rather than about the models.

The T-Rex work from UC Berkeley and collaborators is the strongest independent statement about the mechanism. Its real-world stack pairs a Dexmate Vega-1 bimanual robot with two Sharpa Wave hands, three ZED cameras, Manus gloves and VIVE trackers for teleoperation, and a manufacturer-supplied differential inverse-kinematics package for retargeting. Per-fingertip observations comprise a deformation depth map and a 6-axis wrench. The authors report a tactile-reactive policy reporting a more than 30% higher average success rate than the strongest baseline across 12 tasks requiring delicate force control and deformable-object handling, and they thank Sharpa for equipment maintenance.[19]

The DexTeleop-0 work at Nanyang Technological University is the clearest account of why a lab chose the hand. Assistant Professor Ziwei Wang's group had previously used 6-degree-of-freedom hands and found them unable to perform the wrist-finger coupling motion that screwing and gear meshing require, and found another hand model physically too large for gearbox workspaces. He names precision, ecosystem (URDF files, deployment toolbox, teleoperation integration), and built-in fingertip tactile sensing as the three criteria, and singles out the abduction-adduction freedom at the metacarpophalangeal joints as the specific extra motion his tasks need. The published results show a first-stage success rate of 97.14% on gear meshing and on peg insertion with the group's tactile force-balance controller, against 62.86% and 74.29% respectively without it. Wang told Sharpa that incorporating tactile feedback lifted precise-insertion success from roughly 10% to over 90%.[22][24]

The FTP-1 collaboration places Sharpa inside the research rather than beside it. Sharpa is a listed affiliation, one of its research staff is a co-author, and the acknowledgements credit Sharpa Pte Ltd with hardware, computation, and the Sharpa North-FTP-1 dataset of roughly 4,000 long-horizon dexterous demonstrations. The reported real-robot numbers are modest and worth quoting for calibration: on the Sharpa North setup, FTP-1 scored 45% on drawing a face on a balloon, 80% and 40% on the two stages of repairing a model hand, and 65% on two-handed cap twisting.[20] Those figures come from a paper Sharpa helped fund and should be read as such.

Sharpa also announced Singapore partnerships in 2026 that are deployment commitments rather than research. JTC Corporation and Sharpa signed a memorandum of understanding on April 28, 2026 to deploy a fleet of North robots and open an innovation lab in Punggol Digital District, with JTC providing facilities and access to its Open Digital Platform. The Economic Development Board's June 2026 account groups that with partnerships announced at Sharpa's inaugural AI Robotics Summit involving A*STAR and Grab, covering food and beverage, retail, and port container handling.[31][32]

## How SharpaWave compares with other dexterous hands

The dexterous-hand field splits into research-standard hands built in Europe, Korea, and the United States, prosthetic hands adapted for robots, and a dense cluster of Chinese suppliers selling hands as components to humanoid integrators. See [Dexterous hand](https://aiwiki.ai/wiki/dexterous_hand) for the field as a whole and [Humanoid robot hands](https://aiwiki.ai/wiki/humanoid_robot_hands) for the per-robot picture.

Every cell below comes from the maker's own published material. Blank price cells mean the maker publishes no price, which is the norm in this market.

| Hand | Maker | Actuated DoF | Total joints | Fingertip tactile sensing | Published price |
|---|---|---|---|---|---|
| Sharpa Wave | Sharpa | 22 | 22 | Camera-based array, more than 1,000 pixels per fingertip, 6-DoF force output[12][13] | None; about USD 50,000 reported by 36Kr[2] |
| [Shadow Dexterous Hand](https://aiwiki.ai/wiki/shadow_dexterous_hand) | Shadow Robot Company (UK) | 20 | 24 (4 underactuated) | Shadow Tactile Fingertip, up to 5 fitted[33] | None[33] |
| [Allegro Hand](https://aiwiki.ai/wiki/allegro_hand) V5 Plus | [Wonik Robotics](https://aiwiki.ai/wiki/wonik_robotics) (Korea) | 16 | Not published | 360-degree omnidirectional pressure-sensitive tactile sensor[35] | None[35] |
| RH56 series | [Inspire Robots](https://aiwiki.ai/wiki/inspire_robotics) (China) | 6 | 12 motor joints | Not stated on the series page[34] | None[34] |
| SVH 5-finger hand | Schunk (Germany) | 9 drives | Not published on the product page | Elastic gripping surfaces; no fingertip tactile array published[37] | None[37] |
| Ability Hand | [PSYONIC](https://aiwiki.ai/wiki/psyonic) (USA) | Not published | Not published | Pressure sensors with vibration feedback to the wearer; 490 g[36] | None[36] |
| [Optimus Gen 3](https://aiwiki.ai/wiki/tesla_optimus_gen_3) hand | Tesla (USA) | Not published | Not published | Not published | None |

Read against that table, the SharpaWave's distinguishing property is not its degree-of-freedom count, which the Shadow hand has been close to since the mid-2000s, but the combination of full actuation at that count, camera-based tactile sensing inside human-sized fingertips, and published cycle-life testing. The Shadow hand reaches 24 joints but underactuates four of them; the Allegro hand, long the academic workhorse, has four fingers and 16 degrees of freedom; the Inspire RH56 series, by far the highest-volume of these, has six active degrees of freedom driven through linear and linkage mechanisms and competes on cost rather than dexterity; Schunk's SVH is driven by nine motors. [Tesla's Optimus program](https://aiwiki.ai/wiki/tesla_optimus) publishes no hand specification at all, so the widely repeated claim that its Gen 3 hand has 22 degrees of freedom per hand with the actuators moved into the forearm rests on patent analyses rather than a datasheet, and should not be treated as a vendor figure.

The design divide that matters is transmission. Most high-count hands, including Shadow's and Tesla's, route tendons from motors placed outside the hand, which keeps the fingers slim but introduces cable stretch, friction, and wear (see [Tendon-driven](https://aiwiki.ai/wiki/tendon_driven)). Sharpa's route, described by Leiphone and by outside researchers as direct motor drive inside the hand, trades that packaging advantage for stiffness and cycle life, and it is the reason observers describe the engineering problem as fitting 22 self-developed motors plus 5 cameras plus 5 torque sensors into an adult hand envelope.[17] Sharpa is one of several Chinese-supply-chain-based entrants in this category alongside [Xynova](https://aiwiki.ai/wiki/xynova), [Linkerbot](https://aiwiki.ai/wiki/linkerbot), [ZWHAND](https://aiwiki.ai/wiki/zwhand), [Wuji](https://aiwiki.ai/wiki/wuji_hand), and tactile specialist [PaXini](https://aiwiki.ai/wiki/paxini_technology), although its Singapore incorporation and its global sales posture set it apart commercially.

## Why did NVIDIA choose Sharpa's hands?

On May 31, 2026, NVIDIA announced the **NVIDIA Isaac GR00T Reference Humanoid Robot**, described by NVIDIA as its first open humanoid robot reference design, and selected Sharpa to supply the hands. The reference robot pairs a [Unitree](https://aiwiki.ai/wiki/unitree) H2 Plus chassis (which stands nearly 6 feet tall, weighs about 150 pounds, and has 31 degrees of freedom) with a pair of SharpaWave tactile five-finger hands that add 22 degrees of freedom, plus NVIDIA [Jetson Thor](https://aiwiki.ai/wiki/jetson_thor) (Jetson AGX Thor T5000, Blackwell GPU) for onboard compute and the open [Isaac GR00T](https://aiwiki.ai/wiki/nvidia_isaac_gr00t) software and models. Together the hands and body give the robot 75 degrees of freedom. NVIDIA said leading research institutions including [Ai2](https://aiwiki.ai/wiki/ai2), ETH Zurich, the Stanford Robotics Center, and UC San Diego's Advanced Robotics and Controls Laboratory would use the design, with Unitree selling the H2 Plus version in late 2026.[11]

The arithmetic in NVIDIA's phrasing is worth spelling out, because the release states the 22 figure once and the 75 figure once without reconciling them: 22 degrees of freedom belong to each hand, so the pair contributes 44, which added to the chassis's 31 gives 75.[11] NVIDIA's own wording uses the two-word product name, "Dual Sharpa Wave tactile five-finger hands."[11]

NVIDIA founder and chief executive Jensen Huang framed the platform in terms of the broader robotics opportunity: "Humanoid robots will bring physical AI to the world's largest industries, opening a multitrillion-dollar economic opportunity. The NVIDIA Isaac GR00T Reference Humanoid Robot gives researchers a single, open platform to make breakthrough discoveries toward general-purpose physical intelligence."[11] For Sharpa, the selection placed its hands on a widely promoted, NVIDIA-backed research platform alongside Unitree, one of the most prominent [China AI](https://aiwiki.ai/wiki/china_ai) humanoid-hardware makers, only about a year and a half after the company was founded.[11]

The relationship predates the reference robot. Sharpa and NVIDIA co-authored the Tacmap tactile-simulation paper in February 2026, Sharpa distributes Tacmap as an Isaac Lab plugin, and NVIDIA's EgoScale work used a 22-degree-of-freedom Sharpa hand as its target embodiment.[23][25][29]

## How is Sharpa structured?

Sharpa is incorporated in Singapore as Sharpa Pte. Ltd. The company employs more than 100 people across its three offices as of early 2026, with most core team members recruited externally rather than transferred from Hesai. The Singapore headquarters supports global corporate functions and partnerships, the Shanghai office hosts the bulk of the company's research, development, and manufacturing activities, and the Mountain View, California, office handles business operations and customer engagement in the United States. Public reporting describes Sharpa's leadership as deliberately maintaining limited media visibility during its first year. One investor, quoted anonymously by KrASIA, called the SharpaWave demo "jaw-dropping" while also noting difficulty obtaining direct meetings or fundraising information, telling KrASIA: "I reached out as early as August, but they weren't keen to meet."[1][3][4][8]

Sharpa has published no funding announcement, investor list, or valuation. What exists in public is reporting, and it does not agree. Leiphone wrote in March 2026 that investors said Sharpa's previous round already valued it at USD 1 billion and that post-CES speculation put it at two to three billion, while noting that Sharpa raised selectively and that its Singapore incorporation let it take money globally.[17] Singapore's Economic Development Board and JTC both describe Sharpa as a unicorn, that is, a company valued above USD 1 billion, without giving a figure.[31][32] Any specific number circulating for Sharpa's raise or valuation should be treated as unconfirmed.

36Kr offered one explanation for the Singapore incorporation that Sharpa itself has not given: Hesai was added to the United States Department of Defense list of Chinese military-linked companies in January 2024, its share price fell, and its legal challenge was decided against it in July 2025. A Singapore-incorporated, separately owned entity is a straightforward response to that exposure for a company selling hardware into United States research labs.[2]

Sharpa's near-term commercial focus is on scientific research customers, including leading global technology companies and top research universities, with longer-term plans to expand into service environments such as retail, hospitality, and food service, and ultimately into household applications. AI investor Fu Sheng, quoted by 36Kr, described the North robot as having accomplished "the most difficult tasks" among current humanoid demonstrations, reflecting the broader industry view that contact-rich fine manipulation is one of the harder unsolved problems in [humanoid robots](https://aiwiki.ai/wiki/humanoid_robots).[2][4][8]

## Limitations and open questions

- **The published dexterity numbers are research results.** Google DeepMind's 32% to 92% success rates on household fine-motor tasks describe a research policy in a lab, not product reliability. A dustpan task that fails two times in three, or a bulb that seats one time in three, cannot be scheduled into a shift without a person standing by. Google DeepMind labels the category challenging in its own caption.[16]
- **The hardest tasks are the ones that fail.** The failure pattern is not random: unscrewing beats screwing by 56 points, and every task that requires establishing a new contact under uncertainty sits at the bottom of the table. That is the exact capability the hand's tactile stack exists to provide.[16]
- **Almost every hard number is Sharpa's own.** Durability, force, and lifetime figures are internal test claims with no third-party certification, and the one independently reported spec disagreement so far is between two Sharpa documents rather than between Sharpa and an outside tester.[5][12][13]
- **No price, and constrained supply.** Sharpa publishes no price and reportedly could not fill orders through late 2025.[2][12] A component whose only public price is a single secondhand figure is difficult for integrators to design around.
- **The whole robot is behind the hand.** North's production version slipped from mid-2026 to end of 2026, no specification sheet exists for it, and its first commercial deployment was still described as launching in August 2026 as of July 31, 2026.[27][30][31]
- **Founder attention is a governance question.** The three founders run a Nasdaq-listed and Hong Kong-listed LiDAR company at the same time, and Leiphone reported that the arrangement was itself the reason Hesai's board kept the venture outside the listed entity.[17]

## ELI5: What is Sharpa, simply?

Sharpa is a company that builds robot hands that can both see and feel. Each fingertip has a tiny camera and more than a thousand pressure sensors, so the hand can pick up a fragile egg without crushing it or grip a heavy tool firmly, the way your own hand does. The company also built a full robot, named North, that used two of these hands to play ping-pong against people. In 2026, NVIDIA picked Sharpa's hands to go on its first open robot for university labs.[1][3][5][11]

Later that year, Google's AI lab put its newest robot brain in charge of a pair of these hands and wrote down how often it managed ordinary chores. Taking a light bulb out worked 9 times in 10. Putting one back in worked about 1 time in 3. That gap is the honest state of the art: the hardware can move like a hand, and getting software to use it well is still the unsolved part.[16]

## See also

- [North](https://aiwiki.ai/wiki/sharpa_north)
- [Humanoid robots](https://aiwiki.ai/wiki/humanoid_robots)
- [Dexterous hand](https://aiwiki.ai/wiki/dexterous_hand)
- [Humanoid robot hands](https://aiwiki.ai/wiki/humanoid_robot_hands)
- [Gemini Robotics 2](https://aiwiki.ai/wiki/gemini_robotics_2)
- [Apollo 2](https://aiwiki.ai/wiki/apollo_2)
- [LiDAR](https://aiwiki.ai/wiki/lidar)
- [Tesla Optimus](https://aiwiki.ai/wiki/tesla_optimus)
- [NVIDIA Isaac GR00T](https://aiwiki.ai/wiki/nvidia_isaac_gr00t)
- [Unitree](https://aiwiki.ai/wiki/unitree)
- [Robotics](https://aiwiki.ai/wiki/robotics)
- [Reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning)
- [Consumer Electronics Show](https://aiwiki.ai/wiki/consumer_electronics_show)

## References

1. ["Singapore-based Sharpa surfaces as robotics venture linked to Hesai founders." KrASIA.](https://kr-asia.com/singapore-based-sharpa-surfaces-as-robotics-venture-linked-to-hesai-founders)
2. ["The Emergence of the Most Mysterious Robotics Company." 36Kr.](https://eu.36kr.com/en/p/3647460169092745)
3. ["About Sharpa." Sharpa.](https://www.sharpa.com/pages/about)
4. ["Hesai Technology's Three Founders Kick Off New Cross-Border Entrepreneurial Journey with Robotics Company Sharpa." 36Kr.](https://eu.36kr.com/en/p/3646493069594504)
5. ["Singapore's Sharpa Wins CES 2026 Innovation Award for Groundbreaking AI Robotic Hand SharpaWave." PR Newswire.](https://www.prnewswire.com/news-releases/singapores-sharpa-wins-ces-2026-innovation-award-for-groundbreak-ai-robotic-hand-sharpawave-302608419.html)
6. ["Sharpa Robotics Begins Shipping Its SharpaWave Hand, Targeting Human-Level Tactile Sensing." Humanoids Daily.](https://www.humanoidsdaily.com/news/sharpa-robotics-begins-shipping-sharpawave-dexterous-hand)
7. ["AI Robotmaker Sharpa Reaches Key Milestone With Mass Production of World's Most Advanced Human-Sized Robotic Hand." PR Newswire.](https://www.prnewswire.com/news-releases/ai-robotmaker-sharpa-reaches-key-milestone-with-mass-production-of-worlds-most-advanced-human-sized-robotic-hand-302643434.html)
8. ["Sharpa showcases autonomous fine-manipulation robot and new AI model at CES 2026." Robotics and Automation News.](https://roboticsandautomationnews.com/2026/01/31/sharpa-showcases-autonomous-fine-manipulation-robot-and-new-ai-model-at-ces-2026/98478/)
9. ["Robot plays high-speed ping-pong with 0.02-second reaction time." Interesting Engineering.](https://interestingengineering.com/ai-robotics/robot-plays-high-speed-ping-pong)
10. ["Sharpa's Robotic Hand is Real and Deft at IROS 2025." RoboHorizon.](https://robohorizon.com/en-us/news/2025/10/sharpas-robotic-hand-is-real-and-deft-at-iros-2025/)
11. ["NVIDIA Announces NVIDIA Isaac GR00T Reference Humanoid Robot for Academic Research." NVIDIA Newsroom, May 31, 2026.](https://nvidianews.nvidia.com/news/nvidia-open-humanoid-robot-reference-design)
12. ["Sharpa Wave." Sharpa product page, accessed July 31, 2026.](https://www.sharpa.com/pages/wave)
13. ["Sharpa Hand User Manual" (Sharpa Wave), Sharpa, PDF edition accessed July 31, 2026.](https://documentationpdf.sharpa.com/api/pdf/wave)
14. ["Sharpa Hand User Manual." Sharpa product documentation portal, accessed July 31, 2026.](https://sharpa-robotics.github.io/sharpa-docs/)
15. ["sharpa-urdf-usd-xml: Assets for Sharpa hardware." Sharpa Robotics on GitHub, accessed July 31, 2026.](https://github.com/sharpa-robotics/sharpa-urdf-usd-xml)
16. [Parada, Carolina. "Gemini Robotics 2 brings whole body intelligence to robots." Google DeepMind, July 30, 2026.](https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/)
17. [Liu Xin. "Uncovering Sharpa: why is a company that is not yet profitable valued nearly as highly as Hesai?" Leiphone (AI Technology Review), March 17, 2026.](https://www.leiphone.com/category/robot/J0X65m9VSpvUyZQv.html)
18. [Aufranc, Jean-Luc. "Sharpa Wave is a high-end dexterous robotic hand with 22 DoF, high-sensitivity dynamic tactile array." CNX Software, June 2, 2026.](https://www.cnx-software.com/2026/06/02/sharpa-wave-high-end-dexterous-robotic-hand-with-22-dof-high-sensitivity-dynamic-tactile-array/)
19. [Niu, Dantong et al. "T-Rex: Tactile-Reactive Dexterous Manipulation." arXiv:2606.17055, June 15, 2026.](https://arxiv.org/abs/2606.17055)
20. [Yuan, Chengbo et al. "FTP-1: A Generalist Foundation Tactile Policy Across Tactile Sensors for Contact-Rich Manipulation." arXiv:2606.13102, June 11, 2026.](https://arxiv.org/abs/2606.13102)
21. [Paliwal, Bhawna et al. "Do as I Do: Dexterous Manipulation Data from Everyday Human Videos." arXiv:2606.19333, June 17, 2026.](https://arxiv.org/abs/2606.19333)
22. [Liu, Haichao et al. "DexTeleop-0: Force-Aware Bimanual Dexterous Teleoperation with Ego-Centric Perception towards Shared Autonomy." arXiv:2606.23431, June 22, 2026.](https://arxiv.org/abs/2606.23431)
23. [Su, Lei et al. "Tacmap: Bridging the Tactile Sim-to-Real Gap via Geometry-Consistent Penetration Depth Map." arXiv:2602.21625, February 25, 2026.](https://arxiv.org/abs/2602.21625)
24. ["97% success on gear meshing: how NTU's DexTeleop-0 framework makes dexterous teleoperation viable at scale." Sharpa customer stories, accessed July 31, 2026.](https://www.sharpa.com/blogs/all-customer-stories/97-success-on-gear-meshing-how-ntus-dexteleop-0-framework-makes-dexterous-teleoperation-viable-at-scale)
25. ["EgoScale: 20,000 Hours of Video Grow a Robot Hand." Sharpa research blog, accessed July 31, 2026.](https://www.sharpa.com/blogs/research/egoscale-20-000-hours-of-video-grow-a-robot-hand-opening-the-scaling-law-era-of-dexterous-manipulation)
26. ["Moving Beyond Specialist Training: SimToolReal Enables Zero-Shot Transfer for General-Purpose Tool Use." Sharpa research blog, accessed July 31, 2026.](https://www.sharpa.com/blogs/research/moving-beyond-specialist-training-simtoolreal-enables-zero-shot-transfer-for-general-purpose-tool-use)
27. ["Sharpa North, a 67 DoF humanoid robot for productive work." Sharpa product page, accessed July 31, 2026.](https://www.sharpa.com/pages/north)
28. ["Sharpa CraftNet." Sharpa, accessed July 31, 2026.](https://www.sharpa.com/pages/craftnet)
29. ["Downloads." Sharpa developer resources, accessed July 31, 2026.](https://www.sharpa.com/pages/downloads)
30. ["Sharpa brings robot-handmade ice cream to Shanghai." Sharpa, accessed July 31, 2026.](https://www.sharpa.com/pages/sharpawave)
31. ["Bringing dexterous robots to the world: Sharpa's founder on why Singapore is the launchpad for the future of embodied AI." Singapore Economic Development Board, June 30, 2026.](https://www.edb.gov.sg/en/business-insights/insights/bringing-dexterous-robots-to-the-world-sharpas-founder-on-why-singapore-is-the-launchpad-for-the-future-of-embodied-ai.html)
32. ["Sharpa and JTC Partner to Accelerate AI Robot Deployment at Punggol Digital District." JTC Corporation, April 28, 2026.](https://www.jtc.gov.sg/about-jtc/news-and-stories/industry-news/sharpa-and-jtc-partner-to-accelerate-ai-robot-deployment-at-punggol-digital-district)
33. ["Dexterous Hand Series." Shadow Robot Company, accessed July 31, 2026.](https://www.shadowrobot.com/dexterous-hand-series/)
34. ["The Dexterous Hands." Beijing Inspire Robots Technology Co., Ltd., accessed July 31, 2026.](https://en.inspire-robots.com/product-category/the-dexterous-hands)
35. ["Allegro Hand." Wonik Robotics, accessed July 31, 2026.](https://www.allegrohand.com/)
36. ["Ability Hand." PSYONIC, accessed July 31, 2026.](https://www.psyonic.io/ability-hand)
37. ["SVH 5-finger servo-electric gripping hand." SCHUNK, accessed July 31, 2026.](https://schunk.com/us/en/gripping-systems/special-gripper/svh/c/PGR_3161)
38. ["SharpaWave." CES 2026 Innovation Awards Honoree listing, Consumer Technology Association, accessed July 31, 2026.](https://www.ces.tech/ces-innovation-awards/2026/sharpawave/)
39. ["Sharpa Wave SDK." Sharpa Robotics on GitHub, accessed July 31, 2026.](https://github.com/sharpa-robotics/sharpa-wave-sdk)