Humanoid Robots in 2026: Unitree’s IPO, China’s Lead and the Race to Make Robots Useful

humanoid robots from Unitree

Research cut-off: 21 August 2026.

Unitree Robotics arrived on Shanghai’s STAR Market on 19 August with the sort of debut that normally ends serious discussion. The company sold 40.45 million new shares at RMB150.80, raised about RMB6.1 billion (US$904 million), opened at RMB1,100 and closed at RMB845. That was a 460% first-day gain and a closing market capitalisation of RMB341.77 billion, close to US$50 billion. The retail tranche was 5,526 times subscribed after final allocation. (Unitree prospectus; SSE/Xinhua Finance first-day data; Associated Press)

The debut captured a market approaching an inflection point. Robot bodies are becoming cheaper and more capable, manufacturing volumes are rising, and embodied-AI models are improving quickly. Investors are assigning value to the prospect that these three curves will converge into a major new industry.

Unitree is already more than a robot-dog company with a photogenic humanoid sideline. Humanoids accounted for 1.88% of its main-business revenue in 2023, 27.68% in 2024 and 51.78% in 2025. Total revenue reached RMB1.70 billion in 2025. The company shipped 5,511 humanoids that year, with 5,215 meeting its criteria for recognition as sales. It reported RMB590.75 million of adjusted attributable profit. Nearly 44% of main-business revenue came from overseas.

The operating business confirms that humanoid robotics is crossing from research category into industrial sector. The valuation prices in a much larger opportunity than today’s revenue and deployments can support, which leaves investors exposed to both genuine technological acceleration and speculative excess.

On 20 August, the morning after the listing, founder Wang Xingxing addressed the World Robot Conference in Beijing. Reuters reported his belief that embodied intelligence is moving towards a “ChatGPT moment”: a discontinuous improvement that would move robots from narrow training towards broad usefulness. Chinese reporting of the speech gives Wang’s benchmark greater precision. A robot placed in an unfamiliar environment should be able to complete about 80% of tasks across about 80% of unfamiliar scenarios using voice or text instructions. He placed that threshold two or three years away on an optimistic development path and five to ten years away if progress is slower. (Global Times; Chinese speech transcript)

Wang’s message was optimistic about direction and uncertain about timing. Public investors have valued Unitree at roughly US$50 billion ahead of the software and generalisation breakthrough that its founder associates with broadly useful robots.

The industry now has to progress through the sequence that creates economic value:

locomotion → intelligence → manipulation → useful work

Cheaper bodies and better models have accelerated the first two stages. Manipulation increasingly determines whether that progress can be converted into useful work.

Unitree’s IPO changes the humanoid robotics story

Unitree’s IPO is a better industry marker than another funding round. Private valuations can rely on small transactions, preferential terms and limited disclosure. A mainland prospectus forces an issuer to define revenue, units, customers, risks, related parties and capacity plans. Public trading then supplies a visible, if volatile, price.

Unitree can manufacture and sell at a scale that most Western humanoid developers have not publicly demonstrated. Its 2025 humanoid shipment count is in the thousands. Lower entry prices have opened demand beyond a handful of automotive pilots. Much of that demand still comes from universities, research institutes, technology companies, entertainment and data collection, creating a large installed base before autonomous industrial work has reached comparable scale.

The Robot Report’s initial analysis captured the issue neatly. It reported that about 70% of Unitree’s humanoids went to universities and research institutions and quoted industry executives questioning durability, repeat business and commercial deployment. The exact segment percentages are harder to reconcile from public filings, but the larger point is supported: a large part of current volume is platform demand, not labour replacement.

Research platforms put hardware in the hands of developers, create data and train a generation of engineers on a common body. NVIDIA’s CUDA business also grew as a platform before every GPU had a profitable AI workload. Unitree’s installed base could play a similar enabling role, although shipment graphs still reveal manufacturing adoption rather than equivalent hours of productive labour.

From robot dogs to a RMB1.7 billion humanoid business

Wang founded the Hangzhou company in 2016 after work on the XDog quadruped. Laikago, AlienGo, A1, Go1, B1, Go2 and B2 established Unitree’s reputation for making legged robots at prices below earlier research machines. The H1, introduced in 2023, extended that motion-control and actuator experience into a full-sized biped. G1 followed in 2024 as a smaller, cheaper developer platform. H2 arrived in 2025, followed by H2 Plus and lower-cost R1 variants.

The official store, checked on 21 August, lists the R1 from US$4,900, the G1 at US$13,500, the H2 at US$29,900 and H2 Plus at US$100,000. Delivered cost can rise with hands, compute, EDU access, batteries, shipping and tax, but these are still disruptive entry prices. A buyer can acquire a bipedal platform for the cost of industrial equipment rather than a bespoke research programme.

Unitree combines low-cost assembly with development across motors, reducers, drives, control, sensing, lidar, hands, energy management and embodied software. It assembles complete robots and key components internally while outsourcing or buying selected non-core processing, PCBA placement, moulding and machining. The disclosed detail supports a vertically integrated model without assigning an exact in-house percentage.

Its manufacturing model combines production to demand with safety stock. Sales run through direct channels, agents and online platforms including Tmall, JD, Shopify and Amazon. SDKs and developer tools support research use, forming an early ecosystem around the hardware. A mature marketplace for deployable industrial skills could extend that platform advantage further.

The product-mix change is the clearest commercial signal:

YearUnitree total revenueHumanoid share of main-business revenueHumanoids recognised as sales
2023RMB159.1m1.88%5
2024RMB392.8m27.68%412
2025RMB1,699.3m51.78%5,215

Source: Unitree prospectus. Revenue share uses main-business revenue; total revenue is shown separately.

In two years, humanoids moved from a marginal product line to the majority of Unitree’s core business. Lower prices appear to have unlocked demand that more expensive platforms could not reach. They have also concentrated the company’s future around a category whose applications, standards and competitive structure are changing quickly.

What the IPO numbers actually tell us

The IPO sold 40,446,434 new shares, 10% of the enlarged share capital. Multiplying by RMB150.80 gives gross proceeds of RMB6.099 billion. Unitree plans to invest in embodied-intelligence models, robot technology and products, manufacturing and working capital. Its proposed manufacturing project could eventually provide annual capacity for 75,000 humanoids and 115,000 quadrupeds.

The project describes future capacity. Current output, shipments and productive installations are separate measures.

At the first-day close, Unitree traded at approximately 201 times its 2025 revenue and 579 times its 2025 adjusted attributable profit by simple arithmetic. Reuters Breakingviews had already described the offer valuation as about 100 times earnings. The shares retraced sharply on their second day. Investors are paying for an option on a much larger future industry rather than valuing the current cash flows alone.

First-quarter 2026 revenue rose 68.49% to RMB422.84 million, while adjusted attributable profit fell 52.55% to RMB40.25 million as costs increased. Lower prices, heavier R&D and sales investment, and growing competition are already changing the margin profile as volumes rise.

The retail frenzy is best read as a mixture of scarcity and narrative. Unitree is a rare profitable listed pure-play in a sector associated with AI, advanced manufacturing and national strategy. Chinese policy explicitly promotes embodied intelligence and real-scene deployment. Domestic investors have limited alternatives offering direct exposure. The stock also arrived after several years in which generative AI rewarded capital placed around a platform shift.

The scarcity value and industrial logic are genuine. The trading multiples show how much future progress has already been capitalised.

The shipment number that refuses to stay simple

The disagreement over Unitree’s shipment count reveals how immature market measurement remains.

Unitree’s prospectus says it shipped 5,511 pure humanoid robots in 2025 and recognised 5,215 as sales. The 296-unit gap existed because dispatched units had not yet met acceptance or revenue-recognition conditions. An earlier public version of Omdia’s market radar estimated about 4,200 Unitree general-purpose embodied-intelligence robots, compared with 5,168 for AgiBot and 13,318 globally.

Unitree challenged that estimate, initially saying more than 5,500 robots had reached end customers and more than 6,500 had rolled off the line. Its SSE inquiry response then explained the methodological problem. Some counts include pure bipeds; others include wheeled dual-arm machines and humanoid-like embodiments. Some count factory roll-off, others shipment or delivery. Unitree’s disclosed 5,511 excludes wheeled dual-arm robots.

The company recalculated the relevant global pool at roughly 14,600, which would give it about 37.6%. Later Associated Press reporting attributed more than 5,000 each for Unitree and AgiBot to Omdia and described a market of about 15,000, suggesting later revisions.

Omdia built a cross-company estimate before full issuer disclosure, while Unitree reported its own operational definition. AgiBot’s portfolio includes wheeled machines as well as bipeds. “Humanoid shipments” has yet to become a standardised metric, so leadership rankings move with the treatment of morphology, factory roll-off, shipment and acceptance. Productive hours remain a further, largely undisclosed measure.

The 2026 humanoid robot market

The market now includes enough serious platforms to compare manufacturing evidence, pricing, commercial status and manipulation hardware. Manufacturers use different definitions for shipments, deliveries, production and orders, so the table keeps targets separate from actuals.

Company / platformCountryPublic priceActual evidenceTarget, order or capacityCommercial statusHands and contact sensing
Unitree G1/H2/R1ChinaR1 US$4,900; G1 US$13,500; H2 US$29,9005,511 pure humanoids shipped in 2025; 5,215 salesIPO project: 75,000/year capacity when rampedProducts sold, mostly research/platform demand; industrial output not disclosedConfiguration-dependent; H2 Plus reference uses tactile Sharpa hands
AgiBot A2/X2/G2ChinaSelected youth editions RMB98,000–198,000Omdia estimated 5,168 in 2025; company says 15,000th robot rolled off June 2026No comparable shipment target disclosedBatch production and factory demonstrations; paid productive fleet undisclosedOmniHand/SkillHand options; some vision-based fingertip sensing
UBTECH Walker S2ChinaNot disclosedCompany reports 1,079 full-sized embodied humanoids sold in 20255,000 annual capacity in 2026; orders over RMB800m reportedAutomotive/electronics batches; autonomy and productive hours undisclosedFive fingers; force-control claims; tactile detail limited
Fourier GR-3ChinaNot disclosedOmdia estimated 300 in 2025Not disclosedResearch/service developmentFive-finger option; sensing configuration undisclosed
Leju KuavoChinaNot disclosedOmdia estimated 500; filings/industry material cite 577 sales under another definitionNot disclosedEducation, research and pilotsHand and sensing options vary
Galbot G1ChinaNot disclosedUnit count not disclosedNot disclosedWheeled dual-arm retail/industrial/medical pilotsDexterous hands; tactile details not disclosed
LimX OliChinaNot disclosedUnit count not disclosedNot disclosedDeveloper platform and industrial developmentOptional hands; detail not disclosed
Deep Robotics DR02ChinaNot disclosedUnit count not disclosedNot disclosedIP66 industrial positioning; installed fleet undisclosedConfiguration-dependent
RobotEra L7/STAR1ChinaNot disclosedUnit count not disclosedNot disclosedResearch, performance and industrial developmentDexterous-hand options; tactile configuration undisclosed
Figure 03USNot disclosedMore than 350 delivered from BotQ by April 2026, including internal/data uses12,000/year design capacity; 100,000 over four yearsBMW work and other commercial developmentFive-finger hands; tactile/force detail not fully public
Tesla OptimusUSNot for saleNo customer shipments disclosedProduction expected later in 2026Internal production line and training academyFive-finger hand shown; tactile detail not disclosed
Apptronik ApolloUSNot disclosedShipment count not disclosedPartner programmes with Mercedes-Benz, GXO and JabilPilots/agreementsFive-finger hands; sensing detail not disclosed
Agility DigitUSReported RaaS about US$30/hourPaid GXO deployment; unit count not disclosedNot disclosedNarrow commercial tote handlingSimple end effectors, not dexterous hands
Boston Dynamics AtlasUSNot disclosedEarly production; count not disclosedHyundai programme through 2028Productised platform and pilotsPurpose-built hands; tactile detail not disclosed
Sanctuary PhoenixCanadaNot disclosedShipment count not disclosedNot disclosedRetail/automotive pilots; teleoperation importantHighly dexterous, sensor-rich hands
1X NEONorway/USUS$20,000 or US$499/monthNo verified customer deliveries by 21 August10,000/year factory capacity; 100,000+ ambitionPre-order/internal testingSoft tendon hands; tactile specification undisclosed
NEURA 4NE-1GermanyReservation; price undisclosedNo verified shipmentsRelease targeted for end-2026Development/reservationIntegrated force/torque sensing; local tactile detail undisclosed

Figure’s “delivered” robots include units used for its own data and development. AgiBot’s 15,000 is a cumulative production milestone across several forms. UBTECH’s RMB800 million figure describes orders. Tesla has given a production expectation, and 1X is taking pre-orders. Direct comparisons require those categories to remain separate.

The clearest paid use is often the least human-looking manipulation. Digit moves totes with simple end effectors. It is commercially relevant precisely because the task has been narrowed enough to work.

China’s hardware advantage and the US lead in AI

China’s early lead is strongest where physical industry matters. Motors, reducers, bearings, batteries, power electronics, machined parts and contract manufacturing are available inside dense supplier networks. Domestic competitors iterate quickly and can test in electronics and automotive factories. Lower prices expand the developer base. Central and local policy brings facilities, procurement, demonstration zones and data programmes.

China’s Ministry of Industry and Information Technology has moved beyond broad strategy to real-scene application programmes, asking regions to organise deployments in production and daily life, collect data and explore insurance mechanisms. AgiBot reached its 15,000th factory roll-off in June, another measure of manufacturing momentum even though customer shipment and productive deployment follow later stages.

The competitive picture varies by layer:

  • Manufacturing and hardware cost: China leads on visible volume, supplier density and entry price.
  • Actuators and components: China has broad capacity, though PwC China’s 2026 analysis still identifies gaps in some high-precision reducers, dexterous hands and specialist chips.
  • AI and compute: the United States retains major advantages in frontier models, GPUs, capital and platform software.
  • Industrial integration and safety: Europe retains deep automation expertise and demanding manufacturing customers, with lower public humanoid volumes.
  • Manipulation: strong laboratory results and pilots exist in all three regions, leaving the general-purpose milestone open.
  • Deployment data: China’s growing installed base can provide an advantage when robots are instrumented and used in tasks that generate suitable training data.

The United States has also made Chinese market access harder. FCC order DA 26-786, released on 28 July, added foreign-produced advanced robotic devices to the Covered List unless conditionally approved by the Department of War. The rule blocks new FCC equipment authorisations for covered future devices while leaving existing authorised robots in place. It applies by place of production rather than making a Unitree-specific finding.

Security concerns and regulatory actions have to be kept distinct from proven misconduct. The proposed GUARD Act remains proposed legislation. Commercially, Chinese OEMs face reduced access to an important research and enterprise market, while US buyers lose some low-cost hardware options.

From shipments to useful robots

Humanoid progress is easiest to see in walking, running, dancing, martial arts and backflips. These feats demonstrate actuator power, balance, control and mechanical robustness. They are legitimate engineering achievements.

Their economic value depends on how those capabilities carry into repeatable work.

A demonstration can be rehearsed, scripted, teleoperated, cut between takes or performed in a mapped environment with selected objects. A worker must arrive repeatedly, deal with variation, recover from error, avoid damage, meet cycle time, and remain available long enough to justify capital and support cost.

The commercially useful questions are less cinematic:

  • What percentage of the task is autonomous?
  • How many human interventions occur per hour?
  • What is mean time between failures?
  • How many productive hours are achieved per shift?
  • What are cycle time, damage rate and first-pass success?
  • How long does a new object or task take to configure?
  • Does the customer reorder after the pilot?
  • What is the fully loaded cost per useful hour?

Public reporting rarely provides these numbers. Orders, partnerships, rollout ceremonies, capacity and successful demonstrations dominate while stable operating metrics are still developing.

A fixed industrial arm, conveyor, autonomous mobile robot or purpose-built machine will often beat a humanoid on cost, speed, safety and uptime in a structured task. Humanoid geometry earns its complexity where mobility, human-compatible environments and task variability create economic value. The return must cover the added cost of legs, hands, batteries, sensing and software.

From general intelligence to the final millimetres of a grasp

Dynamic locomotion has improved rapidly across serious platforms. Falls, battery endurance, rough terrain and long-duration reliability still require engineering, but the ability to manipulate variable objects is becoming more commercially differentiating.

Contact changes the state of the world, which makes manipulation a different class of problem.

Before contact, a camera can estimate object identity, geometry and pose. After contact, the hand may hide the object. A bag deforms. A cable flexes. A bottle rotates. A ripe fruit bruises. A smooth package begins to slip. The robot must coordinate fingers, wrist, elbow, shoulder, torso and sometimes both hands while the contact patch changes faster than a high-level vision model should control it.

A robot that can pick unfamiliar products thousands of times without dropping, crushing, misaligning or damaging them may create more economic value than one with exceptional running speed.

Wang Xingxing made this problem unusually concrete at the World Robot Conference. In the Chinese transcript of his 20 August speech, he said today’s models can often understand a task and generate broadly correct motion. Success then breaks down in the last few centimetres or millimetres. A hand appears to have almost grasped the object, but residual position error and the final tactile feedback are not corrected well enough; the accumulated error can cause the task-success rate to fall sharply. Global Times reporting from the conference records the same example.

Wang was describing the gap between an AI model’s input and output and the physical world rather than prescribing a particular sensor architecture. His example still identifies the contact interface as a critical failure point. Recognising an object and deciding to pick it up are perception and planning problems. Correcting the last millimetres after the fingers make contact requires information about what physically happened.

Machine tending, connector insertion, assembly, food handling, sorting, packing, laundry and tool use are contact-rich. They require more than reaching the right position. They require controlling force through uncertain materials and friction. Humans do this partly through touch and fast local reflexes. Robots often try to infer it from cameras, joint torque or wrist force.

Fixtures, compliant grippers, suction, conservative motion and controlled lighting solve many constrained tasks efficiently. General-purpose systems encounter more variation and have fewer environmental constraints to rely on, increasing the value of direct contact information.

As the task distribution broadens, contact information becomes more valuable.

What vision, force and touch each contribute

Vision locates objects, estimates pose and geometry, recognises semantics, plans a collision-free reach and monitors the wider scene. Touch adds information at and after the interface, particularly where fingers occlude the view or the object’s friction, deformation and grip stability cannot be inferred reliably from appearance.

ModalityWhat it tells the robotWhere it is strongTypical limitation
Vision / depthObject identity, pose, geometry, motion and scene contextPlanning before contact; large field of viewContact may be occluded; appearance does not fully reveal friction, grip stability or internal deformation
ProprioceptionJoint position, velocity, estimated torque and body configurationWhole-body control and balanceContact is indirect and often poorly localised
Wrist/ankle force-torqueAggregate three-axis force and torqueCollision detection, insertion and force controlLow spatial resolution; multiple fingertip contacts are combined
Motor current / joint torqueLoad inferred through the drivetrainNo extra fingertip hardwareFriction, transmission compliance and model error reduce precision
Local tactile sensingContact patch, normal/shear distribution, vibration, deformation and sometimes incipient slipGrip correction and contact-rich manipulationCost, wiring, calibration, bandwidth, durability and integration

Normal force acts roughly perpendicular to the contact surface. Shear acts along it. Incipient slip is the local movement that begins before an object visibly falls. Contact location and patch shape reveal whether a grasp is centred or changing. Vibration can help identify texture or a slipping interface. Compliance and deformation indicate how the object or sensor skin is responding.

Tactile sensors measure different subsets of these quantities. A wrist force-torque sensor may provide an excellent total load while telling little about which fingertip is losing grip. A pressure array may measure normal force without shear. A camera-based tactile fingertip can produce dense surface geometry at the cost of more volume and processing. A meaningful specification needs to identify the signal, location, bandwidth and integration.

The emerging architecture is multimodal: cameras for the scene, proprioception for the body, force/torque for aggregate interaction and tactile sensing where local contact state affects success.

As explored in Why Better AI Makes Robot Touch More Important Than Ever, more capable models allow robots to attempt harder tasks. Those tasks create more contact-rich situations and raise the value of better physical observations.

NVIDIA’s tactile Unitree platform points in the same direction

NVIDIA supplied an unusually clear independent signal in May. Its Isaac GR00T Reference Humanoid Robot combines a Unitree H2 Plus body, Jetson AGX Thor compute, Isaac GR00T software, multiview vision and two Sharpa Wave tactile five-finger hands.

The configuration is specific:

  • 31 degrees of freedom in the Unitree body;
  • 22 degrees of freedom in each hand, 75 across body and hands;
  • a stereo head camera with 140° horizontal and 102° vertical field of view;
  • wrist cameras for close manipulation and an IMU for motion;
  • Jetson AGX Thor T5000 with 128GB unified memory and 2,070 FP4 TFLOPS;
  • Isaac Teleop for demonstrations, GR00T models, Isaac Sim and Isaac Lab for training/evaluation, and Isaac ROS for deployment;
  • late-2026 availability.

Sharpa says each fingertip contains more than 1,000 tactile pixels. Supplier pages describe different sensitivity configurations across Wave versions.

The reference system for frontier academic research combines multiview vision with highly articulated tactile hands. NVIDIA’s own description links dexterous manipulation, sensing, control and onboard AI. Parallel grippers remain the efficient choice for many warehouse and industrial tasks, while the five-finger design gives researchers a wider manipulation envelope.

NVIDIA has selected AI + vision + proprioception + force/torque control + local tactile information as the development stack for sophisticated manipulation on its general Physical AI research platform.

Wang identifies generalisation and final contact correction as central obstacles. NVIDIA’s reference platform places tactile dexterous hands on Unitree hardware. Recent research reports better performance on contact-rich tasks when tactile information enters the control and learning loop. These independent signals converge on the same engineering requirement: intelligence has to remain connected to physical feedback at the point of action.

NVIDIA’s broader strategy reinforces the direction. Cosmos world models, Isaac Sim/Lab, synthetic data, teleoperation and GR00T connect data generation, training, evaluation and on-robot inference. Robotics teams need an end-to-end loop that can capture physical interaction, learn a policy, test it and deploy it safely.

Does touch actually improve robot performance?

Recent scientific work is beginning to quantify what touch contributes, primarily in controlled laboratory tasks rather than long-duration industrial operation.

TouchWorld, a July 2026 preprint, reports a hierarchical tactile system that combines slower prediction with faster reaction. Across six long-horizon laboratory tasks, the authors report 65.0% success in clean tests and 53.7% under human perturbation, improvements of 15.7 and 18.5 percentage points over their strongest selected baseline. The tasks include error recovery and changing contact; the results remain specific to an unreviewed preprint and its experimental setup.

FORTE integrates pneumatic sensing into a compliant gripper. Its authors report average force error of 0.2N over a 0–8N range, slip detection within 100 milliseconds and 92% success handling selected fragile, slippery and deformable objects including raspberries and crisps. The experiment illustrates the mechanism under a defined set of tasks.

A GelStereo study published in IEEE Transactions on Neural Networks and Learning Systems reported 95.79% slip classification and demonstrated closed-loop grasping and screwing. A 2026 neuromorphic visuotactile paper reported high-accuracy, low-latency slip detection in selected bottle-cap tasks.

Together, the results show several commercially relevant effects:

  1. local contact signals can reveal a failure before it becomes visible as a dropped object;
  2. force and slip feedback can close a fast corrective loop;
  3. tactile observations can help learned policies distinguish physically different states that look similar to cameras.

Suction, fixtures, compliant materials or slower motion remain effective alternatives for many applications. Wider adoption of tactile sensing will depend on surviving abrasion, impact, contamination and repeated cycles while delivering calibrated, time-synchronised signals to the control stack.

Commercial value appears when the complete system reduces failed picks, damage, intervention or training time at acceptable cost.

Touch as a Physical AI data layer

Tactile sensing can also become part of the training record, extending its role beyond real-time control.

A useful robot trajectory can contain RGB and depth images, language/task context, actions, joint positions, motor currents, torque estimates, wrist forces, local contact, shear, slip, vibrations and success or failure. That sequence gives a learning system more than an outcome. It provides clues about why the outcome occurred.

Unitree makes this distinction in its own SSE response. Autonomous-vehicle data mainly records a system trying to avoid contact. Robot data contains active, controlled contact: force, touch, joint torque, foot pressure and dual-arm coordination. That is a different data problem.

For imitation learning, a tactile trace can show how a demonstrator increased grip during an incipient slip. For reinforcement learning, it can make failure states observable before a binary drop. For grasp prediction, it can connect visual appearance to friction and deformation. For vision-language-action models, it offers a path from semantic intent to the local physics of execution.

RoboTacDex is one example of this direction, assembling visual-tactile-action data for dual-arm dexterous humanoid manipulation. GenForce attacks a related problem: transferring force calibration across different tactile hardware.

Compared with internet images and text, tactile datasets are small. Sensors differ in geometry, units, sampling rate and drift, and physical rollouts make contact data expensive. Cross-sensor transfer remains an active research problem because a policy trained on one fingertip may encounter a different signal distribution on another. Simulation helps, although friction, soft materials and contact transitions are demanding to model.

Tactile data has the potential to become a valuable Physical AI data stream. Repeatable hardware, calibration, metadata and policies that show measurable gains will determine whether it also becomes a defensible commercial asset.

Choosing among tactile architectures

Robotic touch covers several technologies solving different parts of the contact problem.

  • GelSight and DIGIT: optical elastomer sensors produce dense images of surface geometry and can infer shear or slip from marker motion. They offer rich spatial information, with trade-offs in optics, volume, compute and elastomer wear.
  • SynTouch BioTac: a fluid-filled biomimetic fingertip combines deformation, pressure, vibration and thermal channels. It offers rich multimodal data in a fingertip-specific package.
  • XELA uSkin: magnetic three-axis taxels measure local normal and shear forces in thin modular patches suitable for fingers, palms and grippers. Density and magnetic/temperature compensation are design considerations.
  • Contactile PapillArray: soft optical pillars measure local three-dimensional force and displacement, from which torque, slip and friction can be inferred. The modular compliant design differs substantially from camera-imaged gels.
  • TacTip: a soft optical biomimetic architecture tracks internal pins to recover contact geometry and deformation. It has a strong research and sim-to-real ecosystem.
  • Capacitive and piezoresistive arrays: potentially thin and scalable, commonly strong in normal-pressure sensing. Shear, hysteresis, drift and cross-talk depend heavily on design.
  • Magnetic soft sensors: compact three-axis response with calibration and interference trade-offs.
  • Six-axis force/torque sensors: mature and valuable at wrists, ankles and tool flanges, but they aggregate contact and do not provide a dense fingertip map.
  • Tactile skins: extend coverage to palms, arms or bodies, creating challenges in wiring, calibration, repair, durability and cost.

Resolution alone offers a poor basis for comparison. A coarse, robust 3D force taxel may be more useful for a gripper than a dense optical image. Fingertips and forearms require different packaging, while a wrist force sensor may solve insertion without any artificial skin.

The human fingertip analogy is useful only up to a point. Commercial robots do not need to recreate biology. They need the least complex sensing architecture that delivers the required reliability.

From the AI model to the fingertip: where KiriSense fits

Wang’s “ChatGPT moment” sets a demanding goal: robots that generalise across unfamiliar tasks and environments from ordinary language instructions. His grasping example identifies where model-level competence currently loses contact with the physical world. The robot understands the instruction, approaches the right object and produces an almost-correct motion. The final correction after contact still determines success.

That last few millimetres of a grasp require a different kind of information from object recognition. The hand needs to detect when contact begins, how force is distributed and whether the object is stable inside the grasp. Normal force indicates how hard the robot is pressing into a surface. Shear describes forces acting laterally across that surface. Changes in shear and contact behaviour can reveal incipient slip, when an object begins to move within the grasp before a gross failure or drop becomes visible.

KiriSense is developing a compact optical kirigami tactile-sensing architecture aimed at extracting information associated with contact, normal force, shear and slip. The architecture is particularly relevant to fingertips and other constrained robotic surfaces, where useful contact information has to be obtained within a limited hardware envelope. A Henry Royce Institute-backed project with the University of Sheffield is developing a shear-sensing fingertip demonstrator.

Wang’s benchmark depends on generalisation, yet his manipulation example fails at the interface between fingers and object. KiriSense is working on information from that interface. Wang framed the issue as physical-model alignment rather than endorsing a particular sensor architecture. Tactile information can contribute by giving the model and controller evidence about what happened after the planned action entered the physical world.

NVIDIA’s reference architecture and the recent tactile-manipulation studies point in the same direction. More capable robot models are being paired with richer contact information because intelligence needs a fast physical feedback loop. KiriSense sits at that boundary between increasingly capable AI and the object being handled. Its strategic relevance can therefore increase as robot intelligence improves and robots attempt a wider range of manipulation tasks.

Potential applications extend across dexterous humanoid hands, warehouse picking, food handling, industrial manipulation, collaborative robots, agricultural robotics, prosthetics and other embodied-AI systems. Each presents different requirements for force range, geometry, durability, integration and control, but all can benefit from earlier knowledge that a contact state is changing.

For a deeper treatment of AI and robot observability, see Why Better AI Makes Robot Touch More Important Than Ever.

Humanoids could accelerate a much broader tactile-sensing market

Public attention concentrates on the company whose name is painted on the robot, while the value stack extends through:

  • actuators, motors, bearings, reducers and roller screws;
  • dexterous hands and task-specific grippers;
  • cameras, encoders, IMUs, lidar, force/torque and tactile sensors;
  • batteries, power electronics and thermal management;
  • edge compute and networking;
  • simulation, teleoperation, synthetic data and model tooling;
  • system integration, safety, fleet operations and maintenance.

TrendForce has identified roller screws, force/torque sensors, hollow-cup motors and other precision components as material shares of representative humanoid bills of materials. Exact percentages vary by architecture, but rising production creates component demand below the OEM layer. Its 2026 outlook forecasts sharp shipment growth and intensifying competition across systems and components, influencing capacity investment throughout the supply chain.

Component suppliers can grow across competing platforms without depending on a single OEM becoming dominant. For KiriSense, potential customers include humanoid manufacturers, dexterous-hand and gripper companies, industrial-robot producers, research platforms, system integrators and specialist manipulation businesses. Unitree’s success can accelerate the category without becoming the only route to market.

Humanoids may be an especially visible catalyst rather than the entire tactile-sensing market. Industrial arms, cobots, warehouse systems, food and agricultural robots, prosthetics and other machines encounter many of the same contact, force, shear and slip problems. A sensing architecture that solves a cross-platform manipulation problem can participate in several growing markets.

Designs and interfaces are still evolving, qualification cycles can be long, and some OEMs will integrate components internally. The strongest supplier positions will come from measurable task improvements and an integration path that works across several hands, grippers and robot types.

What to watch from 2027 to 2030

The approaching inflection will become visible in operating evidence as well as engineering milestones.

Hardware development will focus on safety, energy efficiency, maintenance and field reliability. Hands need dexterity at acceptable cost and durability. Sensors have to survive contamination and repeated impacts. Battery swaps, charging, fleet orchestration and fall recovery need routine procedures. More stable component interfaces would allow suppliers to scale across platforms.

The intelligence stack is moving towards stronger generalisation and faster recovery. A useful system should learn a new object without thousands of bespoke demonstrations, detect when the world has diverged from its plan, and request help before causing damage. Teleoperation will remain valuable for data and exception handling; intervention frequency will show how autonomy is progressing.

Manipulation brings scene perception, grasp selection, compliant control, local contact feedback, failure prediction and recovery into one reliability stack. Tactile sensing will create the most value where local contact state materially changes task success, with the final architecture shaped by task and cost.

Commercially, customers need a cost per useful hour. Purchase price is only the start. Integrators must include engineering, supervision, floor changes, safety, maintenance, downtime and support. A cheaper robot that needs constant intervention can cost more than an expensive specialised machine.

The most revealing metrics between 2027 and 2030 will be:

  1. autonomous productive hours per robot per month;
  2. interventions and safety stops per hour;
  3. first-pass task success and recovery success;
  4. cycle time relative to a person or specialised machine;
  5. object/part damage and mispick rates;
  6. uptime, mean time between failures and field-service cost;
  7. time and data required to learn a new task;
  8. repeat orders after pilots;
  9. gross margin after support and warranty;
  10. cost per useful hour alongside cost per body.

Customer mix will be equally revealing. Growth in repeat industrial orders, accompanied by customer-disclosed output, would show the transition from platform demand into useful work.

Conclusion: the physical-AI inflection point

Unitree’s IPO shows that humanoid robotics has acquired the capital, manufacturing scale and competitive ecosystem of a genuine industry. Audited revenue, public equity, factories, component supply chains, product families, exports and price competition now sit behind the demonstrations.

China has established an early lead in hardware cost, iteration and visible volume. The United States remains powerful in models, compute and capital. Europe brings industrial integration and automation expertise. The next competitive milestone is general-purpose work at scale.

Wang Xingxing believes the industry is moving towards a robotics equivalent of the ChatGPT breakthrough. His 80% benchmark describes a machine that can enter unfamiliar environments, understand ordinary instructions and complete most of the work it encounters. The timing is uncertain, but the direction is ambitious and clear.

Reaching that threshold requires more than better language or vision models. The robot has to close the loop between deciding what to do and understanding what happened when its hand touched the object. Wang’s final-millimetres example, NVIDIA’s tactile Unitree reference platform and the emerging tactile-manipulation research all locate a critical part of the challenge at the contact interface.

The progression is becoming clearer:

better bodies → better models → better manipulation → useful physical intelligence

Tactile sensing sits directly in that progression, supplying information about contact, normal force, shear and slip when a planned action becomes a physical interaction. KiriSense is developing technology aimed at that interface.

The scale of the opportunity extends beyond any single robot company. As more machines acquire the intelligence to attempt variable physical work, the ability to feel and correct contact becomes part of the infrastructure of Physical AI. The commercial winners will turn that closed loop between perception, intelligence, touch and action into dependable work across factories, warehouses, farms, homes and other environments built for people.

For the November update, see humanoid robotics in 2026: technology, financing and the economics of useful work, including Agility’s SPAC financing and a comparison of humanoid, human and conventional automation costs.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top