Nomura talks about humanoid robots: Data is the key bottleneck and core moat, dexterous hands determine the commercialization process.

```

Data collection is replacing hardware as the core constraint for the industrialization of humanoid robots, and the maturity of dexterous hand technology directly determines the commercialization timetable.

According to Wind Trading Desk, Nomura Securities pointed out in its latest China Robotics Industry Report released on July 5 that data has become the "key component" for large-scale humanoid robot deployment. Figure AI’s CEO confirmed this: "The biggest obstacle stopping us from moving from the current stage to large-scale deployment is data. We need massive amounts of data." Nomura estimates that in a scenario with an annual shipment of about 100,000 units, the industry’s annual data demand is about 10 million hours.

The report also notes that among four main data types, real machine tele-operation (tele-op) data, priced at approximately 500-1000 RMB per hour, constitutes the highest value submarket, with a scale of about 2.2-2.5 billion RMB, while simulation/synthetic data, although with the lowest cost, cannot solely replace real machine data. "Closed loop solutions" that cover the entire chain of data collection, transmission, evaluation, training, deployment, and debugging will be the most defensible business model for pure data service providers.

Four types of data show volume-price differentiation; tele-op data has the highest value

Nomura divides humanoid robot training data into four levels, each with significant differences in price and volume, outlining the competitive landscape of data suppliers.

The first level is non-physical data, including first-person (Egocentric/Ego) video and Universal Manipulation Interface (UMI) data, accounting for 40%-50% of total hours but priced at only about 100-300 RMB per hour, corresponding to an addressable market of about 1-1.5 billion RMB by 2026.

The second level is real machine tele-operation data, accounting for about 30% of total hours, priced at about 500-1000 RMB per hour, and a submarket size of about 2.2-2.5 billion RMB, making it the highest priced and most valuable data category.

The third level is fault recovery data, priced at about 400-500 RMB per hour (according to industry research), but since most manufacturers have not completed the deployment feedback closed loop, this data currently remains in low single-digit percentage levels.

The fourth level is simulation/synthetic data, with the lowest cost at about 50 RMB per 10,000 frames and a market size of about 500-600 million RMB.

Tele-operation and fault recovery data are currently the most scarce and highest margin levels, while Ego/UMI data is the fastest growing pool. This "bottom: low-cost synthetic, top: scarce real machine" hierarchical price structure will determine which suppliers can build lasting competitive moats.

Closed loop solutions are the most defensible business model

Full coverage of data collection, transmission, evaluation, training, deployment, and debugging with integrated software and hardware closed loops is structurally the most defensible business model for pure data service providers.

Data-as-a-Service (DaaS) can quickly be monetized by hourly or project fees, but as client data volumes expand, providers lacking evaluation or "brain-level" capabilities face risk of vertical integration by downstream humanoid robot OEMs. The closed loop model can continuously accumulate first-party scenario data, fault samples, evaluation outputs, and deployment telemetry, which is the prerequisite for building a truly data-enhanced cycle and recurring income.

On the boundaries of simulation data utility, the public disclosures of Physical Intelligence, NVIDIA (NVDA US, unrated), and Lightwheel all point to the same conclusion: simulation data is a "force multiplier" for real machine data, not a substitute. Specifically, π0.5 achieves about 94% success rate on multi-step household subtasks, and 75-80% on long-cycle household tasks; NVIDIA's synthetic motion pipeline boosts GR00T N1 real machine performance by about 40% compared to pure real machine training; Lightwheel reports that a synthetic-to-real training ratio of about 10:1 can bring an average model performance improvement of about 30%, raising task success rate from 60% to 85%.

Industrial scenarios (handling, sorting, machine tool monitoring, assembly) are expected to achieve qualitative breakthroughs in 2027-2028, with humanoid robot shipments rising significantly during this period. Large-scale household deployments may not happen until after 2030, with hotel and serviced apartment cleaning likely the early entry point.

The contradiction between dexterous hand size and sensors restricts commercialization progress

Precision assembly and contact-intensive tasks are hard to be covered by simulation, and home deployment after 2030 fundamentally trace back to technical bottlenecks in dexterous hands.

The dexterous hand market faces an unresolved core contradiction: the closer the hand size is to that of a human, the more accurate the mapping between training data collection and downstream operation; but shrinking the form factor leaves insufficient internal space for sensor payloads. Research shows only one domestic manufacturer is considered to have truly reached human hand size, while mainstream tactile-guided dexterous hands and other high degree-of-freedom designs are still clearly too large, weakening the data-execution consistency.

Tactile technology itself also has a ceiling: point pressure sensors cannot sense lateral force or sliding, current electronic skin has poor fidelity on lateral force curves, and even leading full-hand solutions only have about 80 pressure points installed.

For the arm side, market differentiation has emerged: harmonic reducer plus torque sensor solutions (such as Luna/Skye series) are drifting toward industrial robotic arms, with limited bionic features, and humanoid robot application scenarios are hard to find (as per industry research). The core judgement is: high-precision arms only solve the intermediate motion link, while sufficiently dexterous hands can compensate for arm precision deficits; thus, the preferred architecture is to omit torque sensors and harmonic reducers, and concentrate capability at the end effector.

Until the gap in hand dexterity and tactile fidelity is bridged, the value pool of real machine tele-operation data—and the structurally protected suppliers mastering closed loop collection of that data—will continue to be safeguarded.

~~~~~~~~~~~~~~~~~~~~~~~~

The above excellent content comes from Wind Trading Desk.

For more detailed interpretation, including real-time insights and frontline research, please join the 【Wind Trading Desk ▪ Annual Membership】

Risk warning and disclaimerThe market carries risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account any individual user's specific investment goals, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular situation. Investment based on this is at your own risk. ```