Nomura Securities' latest report on China's robotics sector highlights data as the central constraint and competitive advantage for humanoid robot industrialization, with the maturity of dexterous hand technology directly setting the pace for commercial rollout.
According to the report, data acquisition is now the core bottleneck for scaling humanoid robots, surpassing hardware limitations. The CEO of Figure AI is cited as stating that the primary hurdle to moving from the current phase to mass deployment is the need for massive amounts of data. Nomura estimates that with an annual shipment scenario of around 100,000 units, the industry's yearly data demand could reach approximately 10 million hours.
Data Categories Show Divergence in Volume and Value
Nomura segments training data into four tiers, each with distinct price points and volumes. The first tier consists of non-physical data, such as egocentric video and Universal Manipulation Interface (UMI) data, accounting for 40-50% of total duration but with a low unit price of about 100-300 RMB per hour.
The second and most valuable tier is teleoperation data from physical robots, representing about 30% of total duration with a unit price of 500-1,000 RMB per hour, corresponding to a sub-market size of 22-25 billion RMB.
The third tier is fault recovery data, priced around 400-500 RMB per hour, though its current share remains low as most manufacturers have not yet closed the deployment feedback loop. The fourth tier is simulation/synthetic data, the lowest cost option at about 50 RMB per 10,000 frames.
Teleoperation and fault recovery data are the most scarce and high-margin categories in the near term, while egocentric/UMI data represents the fastest-growing volume pool. This layered pricing structure, with cheap synthetic data at the bottom and scarce real-machine data at the top, will determine which suppliers can build lasting competitive moats.
Closed-Loop Solutions Offer the Most Defensible Business Model
A closed-loop solution covering the entire chain of data collection, transmission, evaluation, training, deployment, and debugging is identified as the most structurally defensible business model for pure-play data service providers.
While a simple Data-as-a-Service (DaaS) model allows for quick monetization, suppliers lacking evaluation capabilities risk vertical integration by downstream humanoid robot OEMs as client data volumes grow. The closed-loop model enables the continuous accumulation of first-party scenario data, fault samples, evaluation outputs, and deployment telemetry, which is a prerequisite for building a genuine data-augmentation cycle and recurring revenue.
On the role of simulation data, disclosures from Physical Intelligence, NVIDIA Corp (NASDAQ: NVDA), and Lightwheel point to a consensus: simulation acts as a "force multiplier" for real-machine data, not a replacement. For instance, synthetic action pipelines have been shown to improve real-robot performance by about 40% compared to pure real-machine training.
Dexterous Hand Constraints Hinder Commercialization Pace
The challenges in precise assembly, contact-intensive tasks, and the delayed timeline for home deployment (post-2030) can be traced back to technical bottlenecks in dexterous hands. A core, unresolved contradiction exists: hands closer to human size allow for more accurate mapping between training data and downstream operations, but the reduced form factor leaves insufficient internal space for sensor payloads.
Tactile technology itself faces limitations; point pressure sensors cannot detect lateral forces or slippage, and current electronic skins have poor fidelity on lateral force curves. Even leading full-hand solutions only incorporate about 80 pressure points.
On the arm side, the market is diverging. Solutions using harmonic drives and torque sensors are drifting towards industrial robotic arms with limited biomimetic characteristics, making them less suitable for many humanoid robot applications. The core assessment is that high-precision arms only solve the intermediate motion link, while sufficiently dexterous hands can compensate for arm precision shortcomings. Therefore, an architectural design that omits torque sensors and harmonic drives, focusing capability on the end-effector, is seen as preferable.
Until gaps in hand dexterity and tactile fidelity are addressed, the value pool for real-machine teleoperation data—and the suppliers mastering its closed-loop collection—will remain structurally protected.
Comments