The accusation that split the room

In September 2026, MechaMind founder Shao Tianlan used a WeChat post to name what many people inside the Chinese embodied AI industry quietly suspected. High-valuation robotics companies were inflating revenue through related-party transactions: co-funding data collection centers with local governments, selling robots into those centers, then buying the collected data back with what looked like training budgets. He named Galaxy General directly in the comments. Galaxy General responded that it would rather focus on technology and commercialization than join a public shouting match.

The Financial Times had already described the mechanics in August. It called the chain "circular financing": a robot company helps a local government build a data center, the center buys the company's robots (booking sales revenue), the center hires people to collect data, and the robot company buys that data back (booking another sale). If the data never enters a model, the cycle is just engineered revenue on a balance sheet.

Key numbers from this cycle:

• 15 USD per hour: average global labor cost of one data collector, before equipment and facilities
• 2-5 hours: usable data produced from a 12-hour collection shift
• 9% to 56%: zero-shot task success on unseen homes, without and with human-behavior pretraining
• 550 million: packages moving through China's logistics network every day

The stakes are huge. In the first half of 2026, Chinese embodied AI startups raised roughly 93.5 billion yuan across 322 rounds, about five times the year-ago total. Money at that scale makes the circular financing question urgent: what does genuine data demand look like, and who is actually buying?

What an hour of robot data actually costs

Real data is expensive. One collector kit runs 2,000 to 3,000 yuan. Labor, at the global average, lands close to 15 USD per hour. The operators I read about in this cycle said a collector working a 12-hour day often delivers only 2 to 5 hours of usable material. The rest is invalid: hands obscured, actions off-protocol, multiple people in frame, equipment idling while someone takes a break.

The reporting from India makes the point concrete. At handicraft worksites, workers sit barefoot in circles and chat while they work. Feet drift into frame constantly, and the model cannot tell hands from feet. The fix sounds trivial: have workers sit apart and cover their feet. Some refused, because it broke the social rhythm of the job. Others kept the device running through breaks because they were paid by the hour. Both behaviors degrade the data.

Even clean footage is raw material, not training data. Annotating a first-person video of someone making egg drop soup means labeling hand positions, tool poses, the timing of each action, and a semantic description of what is happening and why. The model does not learn "a cooking video". It learns the logic of actions and object relations. That annotation layer is where most of the cost actually sits.

What gets reported	What it hides	What actually matters
Hours collected	A 12-hour shift yields 2-5 usable hours	Usable hours after filtering and annotation
Workstation count and seats	Stations can run while producing duplicated, off-protocol footage	Usable data ratio and scene coverage
Trajectory count	1,000 trajectories can all cover the same tabletop grasp	Diversity across scenes, objects, and failure-recovery sequences
Equipment sales to data centers	Revenue can come from the same loop that buys robots	Independent buyers, reorder rate, task-success delta

Everyone I followed in this cycle converged on one line: the data is only worth what it teaches the model, and most collection metrics never touch that question.

Figure: pretrain on humans, not robots

Figure attacked the data problem from the opposite end. Instead of renting human collectors in controlled environments, it built Index, a dataset of human behavior, and pretrained Helix 2.5 on it. The company then rented 30 homes across the San Francisco Bay Area, sent its Figure 03 robots into houses they had never seen, and asked them to do three tasks: tidy a living room, make a bed, fold towels. No new data collection. No per-home fine-tuning. One fixed model checkpoint.

The evaluation was strict. Full-task success only, no partial credit. All 13 to 15 toys had to reach the basket. The towel had to be folded and placed. A single human safety intervention failed the run. Under those rules, zero-shot success went from 9% to 56%, roughly a 6x jump. Same downstream task data, same architecture, same optimization. The only variable was whether the model had seen Index pretraining. Helix 2.5 started from random weights entirely on Index, unlike Helix 02, which initialized from an off-the-shelf pretrained vision-language model.

The numbers hold up under scrutiny. No single evaluation task made up more than 1.90% of Index's pretraining data, so the model was not memorizing the test. Index grows so fast that it records about 35 minutes of new human behavior for every second that passes, and Figure has committed 3.5 billion USD in compute to training Helix. In a scaling study, doubling Index data produced smooth, predictable loss improvements, to the point that small-scale runs predicted the 8x-scale result within 0.54% error.

The 9% to 56% gap is the best public evidence that broad human behavior data can substitute for environment-specific robot data. It comes with a serious catch: it needs serious compute. Most startups cannot commit billions to training runs. Figure's edge is exactly that funding gap.

JD Logistics: data as a byproduct of work

JD Logistics does not collect data on purpose. Its 600,000 delivery workers, 36 million square meters of warehouse space, and 70,000 vehicles generate it as a byproduct of operations. China moves 550 million packages a day, about 6,000 every second, and JD's network feels that pressure directly. Every sortation error is a missed service-level target, which gives the company a business reason to care about the long tail of non-standard operations.

The robots are unglamorous on purpose. The 智狼 system lifts bins along racks up to 12 meters high, while most goods-to-person products stop at 3 to 5 meters. JD calculated the payback at two years, versus the industry-standard three-year expectation. The 仓狼 drives into existing manual warehouses, maps itself, and picks at 80 items per hour with 99.9% accuracy, roughly a third of a top human picker but deployable in two weeks without touching the racks. The 异狼 arm has handled 13 million real packages, cut per-unit cost more than 20%, and now runs around the clock with human intervention below 0.1%.

Two things matter for the data debate. First, JD's data comes from the actual failure modes of running a logistics network: deformed cartons, soft bags, peak-day surges. That is exactly what generic collection cannot reproduce. Second, JD is pushing some of it outward. EgoLive, the first open dataset from its 10-million-hour collection program, already has applicants from more than 8 countries and over 100 universities and labs.

The difference between a working loop and a circular one shows up in where the feedback flows.

The simulation middle path

Between human collection and operational byproduct, open simulation data is quietly filling a gap. WorldSimReady-Home, a dataset by Yootta and Shangpin Home Collection, offers 500 household scenes with 31,711 placed objects across 4,462 rooms. Each scene averages 73 m², roughly a small apartment. The assets are CAD-based, then enriched with physical properties most scene datasets skip: mass, density, static and dynamic friction, restitution.

The scale reads as benchmark material, not pretraining material. 1,000 recorded trajectories at 640×480 RGB-D over 12.18 hours works out to about 44 seconds per episode, short tabletop manipulation tasks with a G1 OmniPicker. Enough to evaluate a policy, not enough to train a foundation model. The license is CC BY-NC-SA 4.0, non-commercial, and access is gated: you share your contact information before downloading. The 681 downloads last month are a real signal that the field is starved for anything with correct physics and manual review.

Route	Who runs it	Scale	Economics	Best for	Hard limit
Human ego data	Figure (Index)	35 minutes of new data per second	Billions in compute	General physical skill pretraining	Compute to digest it
Operational byproduct	JD Logistics	10M hours target, EgoLive open set	Data falls out of business	Logistics and warehouse long tail	Tied to one domain
Sim-ready scenes	WorldSimReady-Home	500 scenes, 1,000 trajectories	Free, gated, non-commercial	Benchmarks, task generation	Physical parameters still need real-world grounding

The car industry already ran this experiment. Autonomous driving needed tens of millions of real road hours before synthetic data became useful, because simulation only covers what reality has already shown it. Embodied AI is at the pre-base stage. A robot that has never registered how a cup slips from wet hands will not learn that from a sim scene that models it perfectly. Simulation stretches real data. It does not replace it. That is why the physical parameters in WorldSimReady-Home matter: without correct mass and friction, the robot grasps perfectly in sim and fails at the first real contact.

What practitioners actually say

The operators doing this work describe a market that looks nothing like the headlines.

"I run a data center. Model companies do not buy in bulk. They order 500 to 1,000 hours around one specific scene or task, run it through training, and only come back if the success rate moved. The second or third order is the real validation. Everything else is just footage."

"I found that the most valuable data is the hardest to license. Stores can make employees sign release forms, but you cannot get every customer walking through to consent. The data closest to a real business is the rarest and the most expensive."

The tactile sensor founder made the consistency argument sharper. "When our gloves recorded the same 300-gram bottle grab as 4N, 5N, and 6N across units, the model learned that 4 to 6 newtons all work. In reality 4N drops the bottle and 6N crushes it. You fix consistency before you scale hours. One inconsistent hour teaches the model a range that does not exist."

The market has started voting on this too. Unitree's shares opened at 1,100 yuan on listing day, well above the 150.80 yuan issue price, because only 7.44% of shares floated, so a small number of trades set the price. By mid-September the stock sat at 470 yuan, down 57% from that opening peak. One buyer I came across put in 400 shares at 925 yuan and watched the position fall almost 50%. H1 revenue grew 48.5% to 1.15 billion yuan, but profit excluding one-off items fell 19%, and R&D spending sat at 8.53% of revenue, far below the 20%+ typical of the peer group. The slide was a judgment on the model story, not the hardware.

The funding split is even blunter. In H1 2026, companies making robot "brains" absorbed more than half of all embodied AI funding, while hardware makers took under 13%. Nearly 50 robotics companies are preparing to list in China. Investors have started asking whether the durable asset is the robot that performs or the data and model that make it useful. That is why circularity accusations matter: they distort the signal the market is trying to read.

Common Pitfalls

• Measure hours, not what the model learned. A 12-hour shift yields 2 to 5 usable hours. If your reporting still leads with "hours collected", the pipeline is not connected to training.
• Treat simulation as physics-free. Scenes without mass, friction, and restitution values produce robots that grasp perfectly in sim and drop everything in reality. Use a set like WorldSimReady-Home that includes them, or measure the parameters yourself. Do not guess.
• Scale sensors before standardizing them. The 4N/5N/6N problem applies to vision too: if two cameras disagree on depth for the same object, one hour of footage encodes the disagreement. Consistent noise scales into confident, wrong behavior.
• Buy generic data hoping it fixes a specific task. Buy 500 to 1,000 hours around a defined scene, evaluate, then reorder. The only data with value is data that moved a model to "good enough".
• Confuse evaluation demos with deployment. Thirty rented homes with a fixed checkpoint is an evaluation. JD's 13 million packages in live operation is a deployment. When someone shows you "it worked", ask how many runs, under what intervention rules, and for how long.

One Thing to Remember

The signal that separates real data demand from circular financing is feedback. Real data loops back into a model, the model changes, the task success rate moves, and the buyer orders again. Circular data loops produce booked revenue and an archive of footage that no model ever touches. When you evaluate a data company, a robot company, or a dataset release, follow the loop and see where it terminates.

The Bottom Line

• If you are building a robot that must work in unfamiliar homes, pretrain on broad human behavior before task-specific data. Figure's control experiment is the clearest public evidence: identical task data yields 9% or 56% depending on whether pretraining existed.
• If you work in logistics or warehousing and cannot spend billions on data, pull your training signal from the operation itself, the way JD does with 智狼 and 异狼, and use open sim sets like WorldSimReady-Home for benchmarking, not for physical grounding.
• If you are evaluating an embodied AI company or data center, hold the line on three numbers: usable data ratio, task-success delta after training, and reorder rate. Ignore device counts and hours collected. One thing to watch: the circular financing pattern is exactly what regulators will probe as the queue of robotics IPOs builds.