Data collection is supplanting hardware as the central constraint for the industrialization of humanoid robots, while the technological maturity of dexterous hands directly dictates the timeline for commercial deployment. According to a recent China robotics industry report published by Nomura on July 5th, data has become a "critical component" for the large-scale deployment of humanoid robots. This view is echoed by the CEO of Figure AI: "The biggest blocker for us going from where we are today to large-scale deployment is data. We need massive amounts of data." Nomura estimates that in a scenario with annual shipments of approximately 100,000 units, the industry's annual data demand could reach around 10 million hours.
The report further notes that among four primary data types, tele-operation data from physical robots, with a unit price of roughly 500 to 1,000 RMB per hour, constitutes the highest-value sub-market, with a scale estimated at 22 to 25 billion RMB. While simulation/synthetic data has the lowest cost, it cannot fully replace physical robot data on its own. An "end-to-end solution" encompassing the entire chain of data collection, transmission, evaluation, training, deployment, and debugging represents the most defensible business model for pure-play data service providers.
Four Data Tiers Show Divergence in Volume and Value; Tele-Operation Data Most Valuable
Nomura categorizes humanoid robot training data into four tiers, each showing significant divergence in price and volume, collectively outlining the competitive landscape for data suppliers. The first tier is non-physical data, including Egocentric (Ego) video and Universal Manipulation Interface (UMI) data, accounting for 40% to 50% of total duration. However, its unit price is only about 100 to 300 RMB per hour, corresponding to an addressable market of 10 to 15 billion RMB in 2026.
The second tier is physical robot tele-operation data, representing about 30% of total duration with a unit price of 500 to 1,000 RMB per hour. This corresponds to a sub-market size of 22 to 25 billion RMB, making it the highest-priced and most valuable data category.
The third tier is failure recovery data, priced at approximately 400 to 500 RMB per hour based on industry research. However, as most manufacturers have yet to establish a closed feedback loop for deployment, this category currently accounts for only a low single-digit percentage.
The fourth tier is simulation/synthetic data, the lowest cost at about 50 RMB per 10,000 frames, corresponding to a market size of 5 to 6 billion RMB.
Tele-operation and failure recovery data are the most scarce and highest-margin tiers in the near term, while Ego/UMI data represents the fastest-growing volume pool. This layered price structure—"cheap synthesis at the bottom, scarce physical data at the top"—will determine which suppliers can build a lasting competitive moat.
End-to-End Solutions as the Most Defensible Model
A closed-loop, integrated hardware and software solution covering the entire process from data collection, transmission, evaluation, training, deployment, to debugging is structurally the most defensible business model for pure-play data service providers. A simple Data-as-a-Service model can achieve monetization quickly via hourly or project-based fees. However, as client data volumes grow, suppliers lacking evaluation capabilities or "brain-level" competencies face the risk of vertical integration by downstream humanoid robot OEMs.
The closed-loop model enables the continuous accumulation of first-party scene data, failure samples, evaluation outputs, and deployment telemetry data. This is a prerequisite for building a genuine data-augmentation cycle and recurring revenue streams.
Regarding the role of simulation data, public disclosures from Physical Intelligence, NVIDIA (NVDA US, not rated), and Lightwheel converge on the same conclusion: simulation acts as a "force multiplier" for physical robot data, not a replacement. Specifically, π0.5 achieved a success rate of about 94% on multi-step household subtasks and 75% to 80% on long-horizon household task categories; NVIDIA's synthetic action pipeline improved the performance of the GR00T N1 physical robot by approximately 40% compared to training solely on physical data; Lightwheel reported that a synthetic-to-physical training ratio of about 10:1 yielded an average model performance improvement of around 30%, boosting task success rates from 60% to 85%.
Industrial scenarios (handling, sorting, machine tending, assembly) are expected to achieve a qualitative breakthrough around 2027-2028, with significant growth in humanoid robot shipments during this period. Large-scale deployment in household settings may have to wait until after 2030, with hotel and serviced apartment cleaning being a potential early entry point.
Size-Sensor Trade-off in Dexterous Hands Constrains Commercialization
The difficulties in simulating precision assembly and contact-intensive tasks, and the postponement of home deployment to post-2030, can both be traced back to technological bottlenecks in dexterous hands. The current dexterous hand market faces a core, unresolved contradiction: the closer the hand morphology is to human hand size, the more accurate the mapping between training data collection and downstream operations. However, reducing the form factor leaves insufficient internal space to accommodate sensor payloads.
Research indicates that among domestic manufacturers, only one is considered to have truly achieved human-hand size. Mainstream tactile-oriented dexterous hands and other high-degree-of-freedom designs remain significantly larger, undermining the consistency between data and execution.
Tactile technology itself also has limitations: point pressure sensors cannot detect lateral forces or slippage, and existing electronic skins have poor fidelity on lateral force curves. Even a leading full-hand solution in terms of size is equipped with only about 80 pressure points.
On the arm side, the market is already diverging: solutions using harmonic drives and torque sensors are drifting towards industrial robotic arms, with limited biomimetic characteristics, making it difficult to find suitable application scenarios for humanoid robots.
A core assessment is that high-precision arms can only solve the intermediate motion segment, 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 physical robot tele-operation data—and the suppliers who master the closed-loop for its collection—will remain structurally protected.
Comments