The photo used in the cover is from https://www.therobotreport.com/bmw-group-deploys-figure-03-humanoid-after-tests-previous-version/.
Physical AI Digest is a weekly briefing produced by Klaudia from Physical AI Company xBerry - a tech company based in Poland building tools at the intersection of Physical AI and operations.
The question that defined Physical AI in 2023 and 2024 was whether the technology worked. The question that defined 2025 was whether it could scale. The question this week is different: who owns the data that makes it all possible?
Stats:
| Value | Description |
|---|---|
| $1B | Figure AI's budget for its gig platform: pay humans to perform physical tasks so robots can learn from watching them |
| 13,361 | UBTECH U1 pre-orders ahead of the September 16 first delivery |
| Sep 16 | First consumer humanoid delivery date: UBTECH U1 ships in 12 days |
| $8.6B | Total raised by humanoid robotics companies in 2026, already 1.8x the full year 2025 |
The Data That Makes the Robot Possible
VLA models - the architecture that now powers every major humanoid platform - learn from demonstration. Instead of programming a robot to pick up a cup, you show it thousands of examples of a human picking up a cup, in different positions, different lighting, different cup shapes. The model learns the mapping from what it sees to what the hand should do. The more demonstrations, the better the model. The more varied the demonstrations, the more robust the model in environments it has never seen before.
Figure AI has been generating this data inside BMW factories since 2025. Figure 03's 99% component placement accuracy across 30,000 X3 vehicles came from exactly this kind of demonstration data accumulated in a real production environment. The problem is that BMW data makes Figure 03 good at BMW. It does not make Figure 03 good at a hospital, a warehouse, a kitchen. Each new environment requires new demonstrations.
The gig platform is Figure's answer. Participants perform physical tasks in their own environments. Every session is recorded and labeled as training data for VLA models. $1 billion allocated to human task demonstrators is the clearest possible statement about where Figure believes the constraint is: not in compute, not in hardware, but in the diversity and volume of demonstrations that make a model generalize.
If Figure builds a data flywheel that spans more environment types than any competitor can access through operational deployments alone, it will have a durable advantage that is not replicable by building more robots. Data flywheels compound. Hardware specs do not.
The most valuable thing a robot company can have in 2026 is not a faster robot or a better model architecture. It is a proprietary dataset of human demonstrations in the environments where the robot will actually operate. Figure AI is spending $1 billion to build that dataset in environments it does not yet have robots in. That is a bet on where the moat will be.
The First Marketplace for Robot Training Data
Kinetic Blocks opened its gated beta on September 1 as the first two-sided marketplace for humanoid training data. On one side: companies training VLA models that need demonstration data across specific task categories. On the other side: owners of that data - robotics labs, companies running gig platforms, research institutions with accumulated datasets they are willing to license.
Before Kinetic Blocks, acquiring training data required bilateral negotiations: identify a data seller, agree on format and quality standards, negotiate a license, close a deal. The process typically took months. Kinetic Blocks replaces that with a standardized marketplace: unified licensing terms, data quality grading, checkout. A company can acquire a data license in days rather than months.
The company is targeting a seed round in Q4 2026. Kinetic Blocks does not create training data, but it removes the infrastructure barrier that was preventing the data economy from functioning at market scale.
September 16: The First Consumer Humanoid Delivers
When UBTECH announced U1 in July with 11,000 pre-orders, the consumer humanoid market was still hypothetical. At 13,361 confirmed pre-orders across price tiers from $16,500 to $136,000, the demand is now a recorded number. The September 16 delivery date is the test of whether that number translates into actual units reaching actual homes.
UWORLD U1 Robot: Price, Models, Specs and Release Date | RoboZaps Blog
The UWORLD U1 starts at RMB 119,800 in China. Compare Lite, Pro and Ultra prices, capabilities, delivery timing and the claims still unverified.
blog.robozaps.com
U1's price range spans from luxury appliance territory ($16,500) to enterprise-edge pricing ($136,000). The $16,500 entry point has a defined buyer demographic: early adopters, tech-forward households, small business owners who can justify the capital against labor cost reduction.
The significance of September 16 is categorical, not numerical. One unit delivered to one home on that date changes what Physical AI is. Before that delivery, consumer humanoids are pre-orders. After it, they are products with owners who will report on whether they work. Those early owner reports will define the consumer humanoid narrative for the next 18 months in ways that no demo video or spec sheet can.
The Market That Is Paying Attention
KraneShares documents $8.6 billion raised by humanoid robotics companies in 2026 alone - already 1.8x the full year 2025, with four months remaining. The capital is concentrating around the companies with the most operational hours: Figure AI at $2.34 billion total raised and a $39 billion valuation, NEURA Robotics with $1.4 billion from Amazon, NVIDIA, and the European Investment Bank.
The companies that win the data race in Physical AI will be very difficult to dislodge, for the same reason that the companies that built the largest LLM training datasets are difficult to dislodge in language AI.
What to Watch Next
- UBTECH U1 September 16 first delivery reports: the first owner reviews from real domestic environments will define the consumer humanoid category
- Figure AI gig platform quality metrics: how Figure measures and enforces demonstration quality across a distributed gig workforce will determine whether the flywheel compounds or degrades
- Kinetic Blocks seed round (Q4 2026): seed capital will signal whether institutional investors believe the data economy for Physical AI has enough immediate buyers
- 1X NEO first delivery reports: if both UBTECH and 1X deliver in September, the consumer market will have its first comparative data point
- Competitive responses to Figure's gig platform: if the data flywheel thesis is correct, every major competitor will need an equivalent data generation strategy - watch for similar announcements in Q4
FAQ
Q: Why is Figure AI paying humans to generate training data instead of using its operational deployments at BMW and other customers?
Operational deployments generate excellent data for the specific tasks and environments where the robot is deployed. BMW data makes Figure 03 good at BMW - it does not transfer cleanly to a hospital, a warehouse, a distribution center, or a home. Each new environment type requires new demonstrations. Figure's gig platform generates data across environment types that Figure does not yet have commercial deployments in - so that when Figure enters a new market segment, it has already trained on data from that environment type. The $1 billion budget is the cost of building a demonstration dataset that spans enough environment types to make a genuinely general-purpose robot possible.
Q: What does Kinetic Blocks actually sell, and who are the buyers?
Kinetic Blocks is a two-sided marketplace: sellers list datasets of physical task demonstrations - recorded and labeled video of humans performing manipulation tasks, navigation tasks, object interaction tasks - and buyers license access to those datasets for training VLA models. Buyers are companies building or fine-tuning robot control models: humanoid manufacturers, foundation model companies, enterprise robotics integrators. Sellers are organizations with demonstration data they are willing to license: research labs, companies running data generation programs, robotics companies with proprietary datasets in task categories they are not competing in. Kinetic Blocks standardizes a process that previously took months of bilateral negotiation.
Q: Is UBTECH U1 actually going to deliver on September 16, and does that delivery validate the consumer humanoid market?
Whether the September 16 delivery happens on schedule is unknowable until it does or does not. Hardware production timelines in humanoid robotics have historically been optimistic. What is known is that UBTECH has 13,361 confirmed pre-orders with deposits, a stated price range from $16,500 to $136,000, and a publicly committed delivery date. If delivery happens and the first owners report reliable operation in real domestic environments, it validates the consumer segment the same way the first iPhone delivery validated the smartphone segment - not because of the volume, but because it proves the category is real. September 16 is a date the market will remember in either case.
Top comments (0)