SamOnRobotics
Analysis

The Data Compilers: Why Raw Robot Data Isn’t Training Data

Hours of footage are not the scarce input. The scarce input is turning messy physical-world signals into data a robot can actually learn from.

dataembodied-aihumanoidschina

TL;DR: People talk about hours of robot data the way they used to talk about parameters in language models — more is better. I don’t think that’s quite right. The scarce input is turning messy physical-world signals into training data a robot can actually use. That’s a compiler problem, not a camera problem. I haven’t visited the companies building this layer. The public record is still worth reading, because this may be the better business than the robot bodies themselves.


The simple version: chat models learned from the internet. There is no internet of robot actions. The industry is trying to build one, and a lot of the attention goes to how many hours get recorded. Recording is the easy part. Making those hours usable is the hard part.

Some context on me. I run consumer brands — Cresimo, Nature’s Hangout and WaxonWare — and our products are manufactured in Chinese factories. So I look at this as an operator and an investor. I hold no robotics positions, and I have no advisory roles or deal talks with anyone in this piece.

The supply problem, in public

ACE Robotics put the constraint bluntly to Computer Weekly in August: “The binding constraint for the humanoid and embodied AI industry today is data scale.” By its count the industry has accumulated only around 100,000 hours of manipulation data, against the tens of millions of hours it says the field needs (Computer Weekly).

Most of that data, the same piece says, comes from teleoperation — a person steering a robot through a task while the system records. ACE called that extremely inefficient, because every hour of footage ties up a robot and a trained operator (Computer Weekly).

That’s a real bottleneck. It is also incomplete. Hours are a volume number. They don’t tell you whether the streams line up, whether failures got recorded, or whether another company’s robot can train on what you captured.

Observational data isn’t operational data

A Shanghai company called ArcheBase puts the distinction this way on its website: “Robots need operational data, not just observational data.” Observational data records what happened — images, trajectories, logs. Operational data, as the company describes it, captures “the actions, states, contact, and failures that make robot learning possible” (ArcheBase).

I want to be clear: I have not visited ArcheBase. I’m reading their public materials, and an April interview their founders gave, because this is the most direct public statement I’ve found of the thesis.

In that April interview with TMTPost (in Chinese; my paraphrase), the founders — Yang Zhexuan, Xu Liangwei and Zhang Jiye — said the scarce thing in the industry is no longer how much raw data you collect. It’s the ability to turn messy physical-world signals into training material a robot can actually use (TMTPost).

They describe a five-step pipeline: quality-check incoming data, align it in time and space, “compile” it into task-level features, make it searchable, then package it for customers (TMTPost).

One example from the interview makes the alignment problem concrete. A camera might record at 30 Hz, an IMU at 500 Hz, and the robot’s joints at yet another rate. Until those streams share one clock and one coordinate system, you can’t really train on them. That’s the compiler. It’s unsexy. It’s also the difference between a pile of sensor logs and a dataset.

Two details from that interview stood out to me as an investor:

  • Four embodied-AI companies — Lingchu, Qiongche, AI² Robotics (智平方) and Zhejiang Humanoid — put in an angel round of tens of millions of RMB and became its first customers (TMTPost). When your investors are also your customers, that’s a demand signal.
  • The plan is a real-robot data-collection factory of more than 10,000 square meters, with 400+ robots across 10+ hardware types (TMTPost). That’s a plan, not a facility I can confirm exists today, and not something I have seen.

(I think of this layer the way I think about our own supply chain. The factory gets the attention, but the spec sheet and the QC process are what decide whether the product works.)

This is also being organized as infrastructure

The coordination isn’t just one startup’s pipeline.

In June, the national-local co-built humanoid robot innovation center in Shanghai’s Zhangjiang area launched what it called China’s first alliance for first-person, real-scene embodied-AI data. Members include ACE Robotics (大晓机器人) and Qiongche. The stated goals: shared collection, labeling, cleaning, ownership and data-exchange standards across factories, city services and warehousing (Shanghai enterprise platform).

The same notice says the center has led or joined work on more than 80 standards, and that seven interconnected humanoid training bases now operate nationwide (Shanghai enterprise platform).

Data standards are boring. They’re also how you get an industry to pool what no single company can afford to collect alone.

Taking a step back, you can see three bets stacked on top of each other. Collect more hours (teleoperation, or recording humans, or simulation). Compile those hours into something a model can train on. Then get enough companies to share a format that the data can move. The first bet is visible. The second and third are where I think the leverage is.

So what for operators and investors

My thought is this:

  • If you’re buying robots, ask about the data before you ask about the body. How many hours of real task data is the model trained on? Were the sensor streams actually aligned, or is that a pile of raw logs? Do they capture failures and recoveries, or just successes? Can the data transfer to a different robot, or only to theirs?
  • If you’re investing, the compiler layer may be the better business than the bodies. A pipeline or a data factory can sell to every robot maker. That’s one reading of why, on the public record, four robot companies backed ArcheBase and became its first customers. I’d want to know whether that demand is repeating, and whether the planned factory is real, before I treated the story as proven.
  • If you’re building, the temptation is to collect first and clean later. The interview I cited argues the other way: the scarce skill is the cleaning.

What would change my mind: if raw teleoperated hours, or simulation, turn out to transfer well enough that the compiler layer is a nice-to-have rather than the bottleneck. I’m unclear, candidly, how much of this is already solved inside the big labs and just isn’t in the public record.

Open question for readers: if you run a warehouse or a plant, would you let a vendor record your people’s work — or your robots’ work — to train models? On what terms?

Disclosure: I hold no positions in any company mentioned. No company reviewed this piece. My own consumer brands are manufactured in China. I have not visited ArcheBase or a specialist data-compiler company. This is a reading of the public record, not investment advice.

Sources & notes

  1. Computer Weekly, “ACE Robotics takes aim at embodied AI’s data problem,” Aug 17, 2026: https://www.computerweekly.com/news/366649472/ACE-Robotics-takes-aim-at-embodied-AIs-data-problem
  2. ArcheBase website (sourced background only; I have not visited): https://archebase.ai/
  3. TMTPost / ZhiKer, Apr 17, 2026, in Chinese (paraphrased; sourced background only). The company has three founders — Yang Zhexuan, Xu Liangwei and Zhang Jiye; the published interview is with Yang and Xu: https://www.tmtpost.com/7956994.html
  4. Shanghai enterprise service platform (Zhangjiang AI Innovation Town), Jun 23, 2026, in Chinese: https://desk.sh.gov.cn/pdfpt/gzdt/20260624/5ec5390cf8cb4ee7829fa5620551517a.html