Humanoid Robot Sharpa D01 Links Touch and Training

Humanoid robot developer Sharpa unveiled a new robot, tactile hand and operator glove at IROS 2026 in Pittsburgh on 28 September (29 September in India). The significant idea is the connection between human demonstration, robot touch and repeated learning. Yet the launch is a product architecture and a set of company specifications, not independent proof that the combined system can perform reliable paid work.

Sharpa, a Singapore-headquartered robotics company focused on dexterous manipulation, announced the three products in a company account of the IROS launch. Its dated release specifies 28 September at 21:29 Eastern Time, which falls on 29 September in India. Independent reports from Humanoids Daily, Black Scarab and 36Kr describe the same three-product debut, while identifying different gaps between the specification sheet and observed performance.

Why this humanoid robot launch is really about touch

Much of the humanoid robot industry measures progress in walking speed, hand joints or the spectacle of a single demonstration. A machine doing useful work, however, must keep controlling an object after contact begins. A cup deforms. A tool slips. A handle catches on a surface. The robot needs to sense these changes and correct its motion before an otherwise successful reach becomes a failed task.

Sharpa’s D01, W02 and AE01 address different points of that loop. D01 moves a robot body and arms. W02 gathers tactile information where the hand meets objects. AE01 lets a person demonstrate a movement and receive contact cues while operating the robot. That arrangement could make a consistent demonstration-to-execution workflow easier to assemble. It does not establish how well the eventual learning system handles new objects, longer shifts or unfamiliar workplaces.

What did Sharpa announce at IROS 2026? It announced a D01 humanoid robot built around two arms and whole-body tactile sensing, a lighter W02 hand with broad touch coverage, and an AE01 glove for teleoperation and training-data capture. The company has released hardware specifications and demonstrations, but has not published an independent benchmark for the three products working together.

What the D01 humanoid robot specification shows

Sharpa says D01 combines its dexterous hands with two seven-degree-of-freedom arms and electronic skin over parts of the body. According to the company, the end effector can reach more than 10.5 metres per second in a maximum-speed condition, positioning repeatability is 0.2 millimetres, and control communication can run at 1,000 hertz. The skin is specified to sample contact at 100 hertz with 0.2-newton force resolution. These are different test dimensions; the company does not say they are achieved simultaneously under one workload.

Humanoids Daily reports that the D01 specification lists 66 total degrees of freedom with hands, about 40 kilograms of body weight excluding a chassis, five kilograms of payload per arm and an NVIDIA Jetson Thor processor. Black Scarab likewise examined the published figures but cautioned that a fast catch demonstrated at a conference says little about endurance, intervention rates or operation around people over a full shift. Neither source presents an independently run D01 performance test.

Whole-body touch is a meaningful design choice. A hand sensor may notice an object sliding between fingers, but it cannot by itself detect a forearm pressing into a shelf or the body brushing a worker. Skin across more of the robot could inform a controller about those contacts. Whether it makes a particular application safe depends on mechanical limits, emergency stops, software response and testing in that application. Tactile coverage is a component, not a safety certificate.

For industrial buyers, the useful question is the task envelope: which objects, payloads, tool geometries and environmental conditions were actually tested together? Maximum speed, minimum resolution and contact sampling are inputs to that assessment. A purchasing decision needs repeatable task completion under stated conditions, not a collage of peak figures.

W02 trades one joint for a lighter hand

The W02 has 21 active degrees of freedom and weighs less than 750 grams, according to Sharpa. Humanoids Daily compares it with the earlier W01 configuration at 22 active degrees of freedom and roughly 1.3 kilograms. A lighter hand can lower the load at the wrist and shoulder, while its smaller size can help it enter spaces built for human hands. Neither change automatically improves grip success; that depends on the objects, controller and task.

One distinctive change is sensing coverage. Sharpa says the fingertip Dynamic Tactile Array measures forces from five millinewtons to 30 newtons at one-millimetre spatial resolution. Electronic skin over fingers and palm has a stated range of 0.1 to 20 newtons and five-millimetre spatial resolution. Those values are manufacturer specifications rather than a public independent laboratory measurement.

36Kr’s original analysis explores the decision to remove one active joint at the base of the little finger. It reports Sharpa’s view that this motion is used less often in common grasping and that eliminating it could simplify the structure. The report also points out the unresolved issue: a motion rarely needed in one test set may matter in another. Comparing W01 and W02 on the same tasks, including failure causes and continuous running time, would show whether the trade-off helps.

Sharpa also advertises a zero-gap grasp for narrow objects and an IP54 rating for the uncovered hand. Humanoids Daily notes a higher IP67 listing when the official protective glove is installed. Those ratings and features deserve practical scrutiny: dust and splash protection differ from the effects of oils, repetitive impacts, abrasive packaging and cleaning chemicals. A buyer should request the exact configuration tested and replacement schedule for its tactile surfaces.

AE01 addresses the expensive data problem

Robots can learn from human demonstrations, but useful demonstrations require more than a video of a moving hand. The system must map finger motion to a robotic hand, record what the robot touched, and identify where a grasp succeeded or failed. Sharpa’s AE01 glove uses 22 encoders to capture operator motion, according to the company. Each fingertip has 256 levels of dynamic vibration feedback, intended to give the operator a cue about contact.

Humanoids Daily notes an important distinction: this is vibration feedback, not a stated mechanism that mechanically resists the operator’s fingers with matching force. Sharpa lists wired data output at 500 hertz with five milliseconds of latency, and wireless output at 150 hertz with 10 milliseconds of latency. These are product specifications; an end-to-end demonstration loop also includes the robot, control software, network and display. The effective response time in an application could differ.

At IROS, 36Kr describes a teleoperated bartending demonstration in which an operator wearing AE01 guided D01 through multiple steps. This illustrates that the hardware can be connected in a live workflow. It does not show how much training data is required before the robot could complete the same sequence autonomously, or how often it could do so without human correction.

The data proposition may ultimately matter more than the public demonstration. If an operator can record repeatable contact-rich examples and label failures, a developer gains material to improve a control policy. But the announced glove alone does not answer data ownership, export formats, software integration, privacy of operator recordings or the cost per useful demonstration. Those questions will determine whether the stack is open enough for outside labs and companies.

What is missing before a buyer can judge the system

The three independent reports converge on a key limitation. Humanoids Daily says all three product pages route buyers to a quote request and that the announcement does not quantify integration savings. Black Scarab lists absent public prices, delivery schedules, production capacity, independent tests and D01 customer results. 36Kr says the effectiveness of adapting earlier models and task experience to the new D01 and W02 still needs corresponding task tests. These are editorial observations about disclosure gaps, not claims that the products cannot work.

A useful assessment would ask Sharpa and its customers for comparable trials: task success rate over hundreds of cycles, object types, payload and speed settings, frequency of human intervention, time to learn a new task, failed-grasp causes, repair intervals and sensor wear. A tactile ablation test would compare the same task with and without contact feedback to show whether the added sensing changes outcomes. Where human operators remain in the loop, labour time must be counted too.

There is also a distinction between company history and this launch. Sharpa’s earlier hand being chosen for a reference robot or an earlier commercial deployment can show the firm has built more than a conference prototype. It cannot validate W02’s new sensing configuration, AE01’s haptic usefulness or D01’s field reliability. Evidence should be attached to the exact generation and configuration being sold.

For teams considering a complete humanoid robot, the purchase process should include system documentation, supported robot-control frameworks, replacement parts, service location and clear data rights. For teams interested only in a hand, W02 must demonstrate compatibility with their existing arm and controller. For teams buying a glove, the relevant output is usable labelled training data and operator performance, not merely an encoder count.

Why the announcement matters to India

Sharpa did not announce an India launch, Indian pricing or local customers in its IROS materials. The India angle is therefore a question for the robotics ecosystem, not a claim of local availability. Indian universities, automation startups and manufacturers evaluating a humanoid robot or dexterous hand increasingly face a build-versus-integrate decision: whether to assemble arms, hands, perception and teleoperation from different vendors, or buy a more coherent stack.

The economics of that choice are hard to read without prices and test data. A lighter end effector may simplify arm design; a teleoperation glove may accelerate collection of demonstrations; integrated touch sensors may reduce custom integration work. But these are plausible benefits, not measured savings for an Indian deployment. Teams should request benchmark tasks tied to their own products and shop-floor conditions before treating vendor specifications as a business case.

The broader robot safety and deployment debate also applies here: contact detection must be part of a complete risk assessment. And as our coverage of physical AI training data explains, the quality and cost of demonstrations can become a bottleneck even when the hardware itself improves.

What happens next

IROS attendees can see Sharpa’s display through 30 September in Pittsburgh, according to the company. The next meaningful disclosures will be firm configurations and availability, repeatable integrated benchmarks, and results from customers or independent laboratories. Demonstrations will remain useful, but they should be accompanied by denominators: how many attempts, under which conditions, with what human assistance.

Sharpa’s launch is best understood as a bet that manipulation improves when motion, touch and teaching are designed together. That is a technically coherent direction for a humanoid robot. Whether the D01, W02 and AE01 make complex work more reliable or cheaper remains an empirical and commercial question that the September announcement does not yet settle.

Frequently asked questions

Is Sharpa D01 available in India?

Sharpa has not announced India availability, local pricing or an Indian D01 customer in its IROS launch materials. Buyers should ask the company directly about regional sales and support.

Does the W02 hand have more joints than W01?

No. Sharpa lists 21 active degrees of freedom for W02, compared with 22 for the earlier W01 configuration. The company emphasises lower mass, smaller size and broader tactile coverage; task-level comparisons are still needed.

Can the AE01 glove make the robot autonomous?

The glove captures operator motion and provides vibration feedback during teleoperation. It can help collect demonstrations, but the launch does not prove that those recordings alone make the robot complete tasks autonomously.

Sources: Sharpa’s original announcement and dated company release; independent reporting and analysis by Humanoids Daily, Black Scarab and 36Kr. All technical performance figures in this article are identified as company-listed specifications, not independently verified performance.