The humanoid robot industry is shifting its focus from building better hardware to controlling the most valuable asset: real-world training data. According to a new report from Nomura Securities, the competitive advantage in robotics will belong to companies that master "closed-loop" data systems, not those that simply manufacture better joints or actuators.

Why Is Training Data the Real Bottleneck for Humanoid Robots?

Figure AI CEO Brett Adcock recently articulated the core challenge facing the entire industry. "The biggest obstacle preventing us from moving from the current stage to large-scale deployment is data. We need massive amounts of data," he stated. This insight aligns with Nomura's analysis, which found that at an annual production scale of 100,000 humanoid units, the industry will require approximately 10 million hours of training data per year.

The most valuable type of training data comes from "real-machine tele-operation," where humans remotely control robotic arms to perform tasks. This data commands a premium price of 500 to 1,000 yuan per hour, roughly equivalent to $2,400 to $4,700 in New Taiwan dollars. At scale, this segment alone represents a sub-market worth approximately $10.5 billion to $12 billion annually.

How Does the Data Market Break Down by Type and Value?

Nomura has mapped a clear hierarchy of training data, ranging from cheap synthetic data generated by computers to extremely scarce real-world recordings. Understanding this stratification reveals why some data suppliers will thrive while others face obsolescence.

• Disembodied Data (40-50% of total hours): First-person video and general interface recordings are the most abundant but cheapest, priced at roughly 100 to 300 yuan per hour, creating a potential market of $7 billion in 2026.
• Real-Machine Tele-Operation Data (30% of total hours): The highest-value segment, commanding 500 to 1,000 yuan per hour and forming the primary competitive battleground among robotics companies.
• Failure Recovery Data (smaller share): Recordings of robots recovering from mistakes cost 400 to 500 yuan per hour but remain scarce because most manufacturers lack complete deployment feedback loops.
• Simulation and Synthetic Data (lowest tier): Computer-generated training data costs only about 50 yuan to produce 10,000 frames, with a market size around $3 billion, but cannot fully replace real-world data.

This stratified structure creates a critical insight: companies that can accumulate and control real-machine data will build lasting competitive advantages, while those relying solely on cheap synthetic data risk being outpaced.

Can Simulation Data Replace Real-World Training?

A common misconception in the robotics industry is that computer simulations can eventually replace the need for real-world data. Research from Physical Intelligence, Nvidia, and Lightwheel definitively refutes this assumption. Simulation data functions as a "force multiplier" for real-machine data, not a substitute.

The evidence is compelling. Physical Intelligence's π0.5 model achieved a 94% success rate on multi-step household tasks, while Nvidia's synthetic motion pipeline improved the real-machine performance of its GR00T N1 model by approximately 40%. Lightwheel's research found that using a roughly 10-to-1 ratio of synthetic to real-machine training data can boost a model's task success rate from 60% to 85%. In other words, simulation works best when paired with real-world data, not as a standalone solution.

Why Are "Closed-Loop" Business Models the Only Defensible Strategy?

Nomura issued a stark warning to companies considering a pure "Data-as-a-Service" model. While selling data directly to robotics manufacturers can generate quick revenue, this approach faces a critical vulnerability. As client data volumes expand, robot makers will increasingly build their own data collection and evaluation capabilities in-house, effectively eliminating the need for external data suppliers.

The only truly defensible business model is a "closed-loop solution" that covers the entire workflow of data collection, transmission, evaluation, training, deployment, and debugging. Companies that control this full pipeline can continuously accumulate first-party scenario data, failure samples, and real-world telemetry, building what Nomura calls a "data flywheel" with recurring revenue streams. This approach mirrors how leading AI labs protect their competitive advantages by controlling both model training and deployment.

When Will Humanoid Robots Actually Be Ready for Factories and Homes?

The timeline for humanoid robot deployment depends heavily on solving the technical challenges of dexterous hands, the robotic equivalent of human fingers. Industrial applications like parts handling, sorting, and assembly are expected to see significant growth in 2027 and 2028, but large-scale deployment for home use will likely wait until after 2030.

The bottleneck is a paradox that Nomura identifies as the "size curse." The closer a robotic hand's design matches human hand proportions, the more accurately training data maps to real-world performance. However, miniaturizing the hand to human size creates insufficient internal space to accommodate adequate sensor payloads. Currently, only one Chinese manufacturer has achieved a truly human-sized hand design, while most high-dexterity solutions remain significantly larger, undermining the consistency between training data and actual execution.

Tactile sensing technology compounds this challenge. Point pressure sensors cannot detect lateral forces or slippage, and existing electronic skin has poor fidelity in measuring lateral force curves. Even the most advanced full-hand solutions carry only about 80 pressure points, far below the sensory richness of human hands.

How Are Companies Like Figure AI Proving Humanoids Can Work in Real Factories?

While data and hardware challenges persist, Figure AI is demonstrating that humanoids can already perform meaningful work in industrial settings. Figure's second-generation robot spent ten months at BMW's Plant Spartanburg in South Carolina, where it inserted sheet-metal parts for welding and supported production of more than 30,000 BMW X3 vehicles.

This real-world deployment has now graduated to a new phase. Figure's third-generation robot, called F.03, recently arrived at Spartanburg to take on parts sequencing in logistics, a task that requires sorting unsorted components into sequencing trolleys that feed the assembly line. The new model includes fingertip sensors capable of detecting forces as small as three grams, palm-mounted cameras, and a soft multi-density exterior designed for safe operation near human workers.

BMW has formalized this commitment by establishing a Centre of Competence for Physical AI in Production and is now piloting humanoids at Plant Leipzig in Germany, signaling plans to scale the technology across its global manufacturing network. This progression from pilot to production deployment underscores that the data and experience gained from real-world operation is accelerating the timeline for humanoid adoption in industrial settings.

The humanoid robot race is ultimately a data race. Companies that can build closed-loop systems to collect, evaluate, and deploy real-machine training data will dominate the next decade of robotics, while those betting solely on hardware innovation or simulation will find themselves outpaced by competitors with superior data moats.