According to an industry report, data acquisition is now the central bottleneck for humanoid robot industrialization, supplanting hardware constraints, while the technological maturity of dexterous hands directly dictates the timeline for commercial deployment.
A recent report on China's robotics sector from Nomura points to data as the "critical component" for large-scale humanoid robot deployment. This view is echoed by the CEO of Figure AI, who stated that the primary obstacle to moving from the current phase to mass deployment is data, highlighting the need for massive datasets. Nomura estimates that under a scenario of approximately 100,000 units shipped annually, the industry's yearly data requirement would be around 10 million hours.
Four Data Tiers Show Divergence in Volume and Value
The report segments humanoid robot training data into four tiers, each showing distinct price and volume characteristics that shape the competitive landscape for data suppliers.
The first tier is entity-free data, including Egocentric/Ego video and Universal Manipulation Interface (UMI) data. This category accounts for 40-50% of total data hours but carries a relatively low unit price of approximately 100-300 RMB per hour, corresponding to an addressable market of about 10-15 billion RMB by 2026.
The second tier is real-machine teleoperation data, representing about 30% of total hours. With a unit price of roughly 500-1,000 RMB per hour, this forms the highest-value sub-market, estimated at 22-25 billion RMB.
The third tier is fault recovery data, priced around 400-500 RMB per hour based on industry research. However, as most manufacturers have not yet established a closed-loop for deployment feedback, this data type currently accounts for only a low single-digit percentage.
The fourth tier is simulation/synthetic data, which is the lowest-cost option at about 50 RMB per 10,000 frames, corresponding to a market size of 5-6 billion RMB.
Teleoperation and fault recovery data are the scarcest and most profitable tiers in the near term, while Ego/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 Strongest Defensive Business Model
A closed-loop solution encompassing the entire process from data collection, transmission, and evaluation to training, deployment, and debugging is structurally the most defensible business model for pure-play data service providers.
While a simple Data-as-a-Service (DaaS) model allows for quick monetization on an hourly or project basis, suppliers lacking evaluation capabilities or "brain-level" competencies face risks of 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 streams.
Regarding the role of simulation data, disclosures from Physical Intelligence, Nvidia (NVDA US, not rated), and Lightwheel converge on the same conclusion: simulation acts as a "force multiplier" for real-machine data, not a replacement. Specifically, π0.5 achieves a success rate of about 94% on multi-step household subtasks and 75-80% on long-horizon household task categories; Nvidia's synthetic action pipeline improved the GR00T N1 real-machine performance by approximately 40% compared to pure real-machine training; Lightwheel reports that a synthetic-to-real training ratio of about 10:1 yields an average model performance improvement of around 30%, boosting task success rates from 60% to 85%.
Industrial scenarios (e.g., handling, sorting, machine tending, assembly) are expected to see a qualitative breakthrough around 2027-2028, with significant growth in humanoid robot shipments during this period. Large-scale deployment in household settings, however, may not occur until after 2030, with hotel and serviced apartment cleaning likely being an earlier entry point.
Dexterous Hand Size-Sensor Dilemma Constrains Commercialization
The fundamental challenges of precise assembly and contact-intensive tasks being difficult to simulate and household deployment being a post-2030 story can be traced back to the technological bottlenecks of dexterous hands.
The current dexterous hand market faces an unresolved core contradiction: the closer the hand morphology is to human hand size, the more accurate the mapping between training data collection and downstream operation. However, reducing the form factor leaves insufficient internal space for sensor payloads. Research indicates that among domestic manufacturers, only one is considered to have truly achieved human-hand size, while mainstream tactile-oriented dexterous hands and other high-degree-of-freedom designs remain significantly larger, weakening the consistency between data and execution.
Tactile technology itself has limitations: point pressure sensors cannot detect lateral forces or slippage, and existing electronic skins have poor fidelity on lateral force curves. Even the leading full-hand solutions in terms of size are equipped with only about 80 pressure points.
On the arm side, the market has begun to diverge. Solutions combining harmonic drives with torque sensors (e.g., Luna/Skye series) are gradually drifting towards industrial robotic arms, offering limited biomimetic characteristics and making it difficult to find suitable application scenarios for humanoid robots, according to industry research. A core insight is that high-precision arms can only address intermediate motion, while sufficiently dexterous hands can compensate for arm precision shortcomings. Therefore, there is a preference for architectural designs that omit torque sensors and harmonic drives, concentrating capabilities on the end-effector.
Until the gaps in hand dexterity and tactile fidelity are bridged, the value pool of real-machine teleoperation data—and the suppliers mastering the closed-loop for collecting it—will remain structurally protected.
Comments