A robot that can fold a shirt in a lab often freezes in a real bedroom, and the reason is usually the same: it has not seen enough of how people actually move. A Beijing company just dumped a very large pile of that missing experience into the public domain. The bet is that open data, not secret data, wins the standards war.
For embodied intelligence (具身智能), data has been the choke point. A language model can scrape the internet; a robot has to learn from physical action, which is slow, expensive and messy to collect. One Beijing startup thinks the way through is to give the data away.
What Guanglun released
The company is Guanglun Intelligence (光轮智能), a Beijing-based "embodied data" firm. On August 20, 2026, during the World Robot Conference in Beijing, it released EgoSuite-Open100K — described as the first open dataset of its kind at the 100,000-hour scale of first-person human behavior.
The shape of the data:
- 100,000 hours of egocentric (first-person) human action.
- Coverage across 7 environment categories and 128 scenario types.
- More than 15,000 collection scenes and more than 15,000 tasks.
- Captured from head and wrist dual views, with hand and full-body pose, semantic labels and depth information.
The dual view is the clever part. A head camera shows what a person intends; a wrist camera shows how the hand actually moves. Together they teach a model both the goal and the grip — the missing link between "want to pick this up" and "here is how a hand does it."
Not just a data dump
Guanglun paired the dataset with two pieces of infrastructure. RoboFinals is a simulation-and-evaluation base built on its own physics engine, meant to test robots at scale without burning through real hardware. RoboStack is a deployment-and-feedback platform that can orchestrate clusters of robots, allow remote human takeover, and pull failure cases back into the next training loop.
The loop is the point: human data teaches, simulation tests the edge, real deployment collects the failures, and those failures feed the next round of collection. That is a continuous-learning system, not a one-time release.
The standards play
The first batch went live on Hugging Face and AtomGit, and the project is being donated to the OpenAtom Open Source Foundation (开放原子开源基金会). Guanglun's co-founder and president, Yang Haibo, has been explicit that the industry's real problem is not a lack of data but a lack of shared format: different devices, labels and time structures that cannot be combined. Open the biggest dataset you can, the logic goes, and you also get to define what "standard" means.
The company also announced a five-year plan to co-build 10 billion hours of embodied-intelligence data with partners across research, modeling and scene providers. That number is an ambition, not a delivered asset — but the direction is clear.
Why this matters beyond one firm
Through 2026, several Chinese robotics and internet firms opened large behavior and operation datasets, a shift from hoarding to sharing. The shared fear is that, while global standards for embodied data have not hardened, the first movers to open data also set the format, the labels and the benchmarks. Whoever defines the textbook writes the exam.
Guanglun's angle is that human behavior is the "teaching encyclopedia" for robots — and that giving it away, far from losing value, builds the market that then pays for curated, higher-quality data later.
Deployment takeaways
According to Li An, chief scientist at BrainNet (脑机网), open datasets shift the competitive fight from who hoards data to who sets the standard.
The honest catch
Open data is not a finished product. A 100,000-hour corpus still has to be cleaned, validated and matched to a specific robot body and task. And "first-person human action" is only one slice of what a working robot needs — it tells you how people move, not how a particular gripper should close. The physics-simulation and real-deployment halves of Guanglun's story are where the value is supposed to compound, and those are earlier and less proven than the dataset itself.
Honest limitations
- The 100,000-hour scale, 7 environment categories, 128 scenarios and 15,000+ scenes and tasks are company-stated and reported by state and industry media; we did not independently audit the corpus.
- The August 20, 2026 release date and World Robot Conference venue are confirmed by multiple outlets, but the dataset's real-world training impact is not yet measurable.
- Hugging Face and AtomGit availability and the OpenAtom donation are reported by the company and media; we did not verify live download access.
- The 10-billion-hour five-year plan is a stated ambition, not delivered data.
- We found no independent benchmark yet showing models trained on EgoSuite-Open100K outperforming alternatives on standardized tasks.
What readers can do now
- If you train embodied-AI models, pull EgoSuite-Open100K from Hugging Face and use it as a baseline before paying for proprietary behavior data.
- If you build robots, pair open human-data corpora with your own sim-to-real loop — the dataset alone will not close the last mile to a working grip.
- If you invest in AI infrastructure, watch which open dataset becomes the de-facto benchmark; the format owner, not the data owner, captures the ecosystem.