Mecka AI Raises $60M to Build Robot Training Data Engine

The hardest problem in robotics right now is not building the robot. It is finding enough data to teach it how to move like a human. Mecka AI just raised $60 million betting that the answer is strapping sensors to actual people and recording the messy, improvised ways humans pick up a cup, climb a stair, or fold a shirt. The company turns everyday human motion into the fuel for robot brains, and investors are treating that raw motion as one of the scarcest resources in physical AI.

What Actually Happened

Mecka AI has raised a total of $60 million across two previously unannounced fundraises: a $25 million Series A closed in November and a $35 million follow-on investment. Framework Ventures led both rounds, with participation from Menlo Ventures, SV Angel, Kindred Ventures, and angel investor Ted Xiao, a former Google DeepMind researcher who works on robot learning. The company kept the funding quiet until now, an unusual choice in a sector where startups normally announce every dollar to signal momentum to talent and rivals.

The core product is Egoverse, a large-scale, first-person interaction dataset that maps how humans naturally manipulate objects and move through space. Mecka collects this data with a mix of custom hardware, including wearable body sensors, and ordinary iPhones, which let the company scale capture without shipping expensive rigs to every contributor. The goal is to record the long tail of physical behavior: not just clean demonstrations, but the hesitations, corrections, and awkward grips that make human motion robust in the real world and that robots consistently fail to reproduce.

The traction numbers explain the investor enthusiasm. Mecka projects an annual run rate of $100 million based on contracts it has already signed, and a cofounder of Framework Ventures described it as the fastest-growing company the firm has ever backed. Founder Gao has been explicit that he does not want Mecka to be a passive data vendor. He wants the company on the frontlines, helping robotics firms integrate the data and train their models directly, which positions Mecka closer to a partner than a supplier.

Why This Matters More Than People Think

Robotics has a data problem that language models never had. Text and images exist in near-infinite supply on the open internet, which is what made large language models possible. Physical interaction data does not exist that way. There is no web-scale archive of how a hand applies exactly the right pressure to a soft fruit or recovers when a grip slips. Every robotics lab building vision-language-action models hits the same wall: the models are hungry for demonstration data, and almost none of it is just lying around to be scraped.

That scarcity is why a data company can command a $100 million run rate before most consumers have ever heard its name. If Egoverse becomes a default training source, Mecka sits upstream of every robot its customers ship, in the same structural position that labeling and dataset companies occupied during the rise of computer vision. Owning the data layer in a gold rush is often more durable than owning any single application, because the application makers all have to come back and buy more as their models scale.

The timing rides a broader wave. The global robotics market reached $38 billion in 2026, a 34% year-over-year jump that ranks among the fastest growth the sector has seen in a decade, and humanoid programs from Tesla, Figure, and a wave of Chinese manufacturers are all racing to deploy. Every one of those programs needs to teach machines dexterity and locomotion, and every one of them is constrained by the same shortage of high-quality human motion data that Mecka is built to supply. The robots are arriving faster than the data to train them.

The demand is not limited to humanoids that make headlines. Industrial automation, warehouse logistics, food preparation, and elder-care robotics all need fine-grained manipulation data, and these sectors are often more willing to pay than research labs because they have concrete tasks and revenue tied to getting them right. A robot that loads a dishwasher in a commercial kitchen or restocks a shelf has to handle thousands of object shapes and grips, and the only way to teach that range cheaply is to learn it from humans who already do it all day. Mecka is selling into a demand curve that gets steeper as deployment moves from demos to paying customers.

The Competitive Landscape

Mecka is not alone in chasing the data bottleneck, and the rival approaches reveal how unsettled the field is. Some companies, like the web-video-trained efforts that have raised hundreds of millions, try to learn from passive footage scraped online, betting that watching humans on video is enough. Others rely on teleoperation, paying workers to puppet robots through tasks, which produces clean data but at painfully slow speed and high cost. Mecka's wearable-plus-iPhone approach is a third path, aiming for the scale of video with the fidelity of direct measurement that pure video cannot capture.

The deep-pocketed in-house programs are both customers and potential competitors. Tesla generates its own Optimus data through teleoperation and simulation, and Figure has built proprietary data pipelines, meaning the largest humanoid players may never need to buy from Mecka at all. The opportunity is the long tail: the dozens of robotics startups and industrial automation firms that cannot afford to build a global motion-capture operation from scratch. For them, buying Egoverse is far cheaper than recreating it, which is exactly the wedge Mecka is exploiting.

The historical parallel is Scale AI in the autonomous-vehicle era. A decade ago, self-driving startups discovered that the binding constraint was not algorithms but labeled sensor data, and Scale grew into a multi-billion-dollar business by becoming the data backbone for an entire industry. Mecka is positioning itself as the Scale AI of physical robotics, betting that the same pattern repeats: when everyone is racing to build the same kind of model, the company that owns the training data quietly becomes indispensable to all of them, regardless of which robot maker ultimately wins.

The Chinese dimension sharpens the race. Manufacturers like Unitree have driven humanoid hardware costs down dramatically, which means the bottleneck for an entire national robotics industry is shifting from building bodies to training brains. That hands a global data supplier like Mecka a potential customer base far larger than the handful of American humanoid startups, but it also raises hard questions about which datasets cross borders and which governments treat human motion data as strategic. A company sitting on the training corpus for physical AI is sitting on something states will eventually care about, the way they now care about advanced chips and model weights.

Hidden Insight: Whoever Owns the Motion Owns the Robots

The non-obvious bet here is about leverage, not data volume. A robot trained primarily on Mecka's dataset inherits the biases, coverage gaps, and strengths of that dataset. If Egoverse becomes the standard corpus for an entire generation of robots, Mecka does not just sell a product; it shapes how a whole class of machines understands physical reality. That is a position with enormous strategic gravity, because switching training data after a model is built is expensive and risky, which locks customers in far more tightly than a typical software vendor relationship.

This is the quiet reason data companies can outlast the application makers they serve. A robotics startup that trains on Egoverse and then tries to switch suppliers has to re-collect or re-license a comparable corpus, revalidate its models, and absorb the performance regressions that come from a different data distribution. That friction is what converts a one-time sale into a recurring dependency. Mecka does not need to win every robot program; it needs enough of them anchored to its data that leaving becomes more expensive than staying, which is the same dynamic that turned cloud infrastructure and payment rails into businesses customers complain about but rarely abandon.

There is also a privacy dimension that the market has barely priced. Capturing how millions of people move, gesture, and manipulate objects through body sensors and phone cameras creates one of the most intimate datasets imaginable. Biometric motion signatures can identify individuals as reliably as a fingerprint, and the same data that teaches a robot to fold laundry could, in the wrong hands, track or profile the humans who generated it. As Egoverse scales, the regulatory and consent questions around physical behavior data will grow alongside it, and Mecka will be writing norms others later have to follow.

The bear case, however, is that Mecka's moat is thinner than the run rate suggests. Critics argue that human motion data, unlike a proprietary model, can be regenerated by anyone willing to spend on sensors and contributors, and the iPhone-based capture method that makes Mecka cheap to scale also makes the approach easy to copy. If Tesla, Figure, and the Chinese humanoid makers all build internal data engines, Mecka could be squeezed into serving only the smaller players, a real but capped market. The $100 million run rate is built on current contracts, and contracts in a hype-driven sector can evaporate as fast as they appear.

The deeper risk is that the entire premise of human-imitation data could be leapfrogged. Some researchers argue robots should learn primarily in simulation, generating synthetic experience at a scale no human capture program can match, and that real human motion will become a fine-tuning detail rather than the foundation. If simulation-first approaches win, Mecka's painstakingly collected physical data becomes a supplement instead of a substrate. The company is making a concentrated bet that real human behavior remains irreplaceable, and the next two years of robot-learning research will decide whether that bet ages like a moat or a museum piece.

What makes the bet genuinely hard to call is that both camps are partly right. Simulation excels at generating volume and rare edge cases, while real human capture excels at the subtle, physically grounded behaviors that simulators still render unconvincingly. The likely outcome is a blend, and the strategic question becomes which layer captures the value. If real human data ends up as the high-margin fine-tuning step that makes simulation-trained robots actually work in the messy real world, Mecka thrives. If it becomes a commodity input dwarfed by synthetic data, the margins compress. The company is wagering that the hardest, most human part of physical intelligence stays scarce, and scarcity is where pricing power lives.

What to Watch Next

In the next 30 days, watch whether Mecka names any of the customers behind its $100 million run rate. Signed contracts with recognizable humanoid or industrial-automation companies would validate that Egoverse is becoming infrastructure rather than an experiment, while continued silence would leave the run-rate claim harder for the market to evaluate. Also watch whether competitors respond with their own data-partnership announcements, a sign the bottleneck thesis is going mainstream.

Over 90 days, the marker is data scale and coverage. Mecka's value rises with the breadth of human behaviors it captures, so look for announcements about contributor counts, new capture modalities beyond body sensors and iPhones, and expansion into industrial or healthcare motion that commands premium pricing. A company in this position lives or dies by how fast its dataset compounds, and the pace of capture is the clearest leading indicator of whether the moat is widening or stalling.

Over 180 days, watch the simulation versus real-data debate play out in published robot-learning results. If the strongest new manipulation models lean heavily on real human demonstration data, Mecka's thesis strengthens and its pricing power grows. If simulation-trained models start matching or beating them, expect investor enthusiasm for human-capture startups to cool quickly. The robotics field is moving fast enough that the question of what robots should learn from will be answered, at least provisionally, within a year.

Watch the funding climate too. Mecka raised quietly across two rounds, but a million run rate in a hot sector tends to invite a large priced-up round, and how that round is valued will signal whether investors see the company as a durable data monopoly or a fast-growing but copyable vendor. A blockbuster valuation would pull more capital into human-capture startups and intensify the race; a flat or down round despite the run rate would suggest the market has decided the moat is shallow. Either way, the next financing is a clearer verdict on the thesis than any press release.

The robots are arriving faster than the data to train them, and Mecka is betting that whoever owns humanity's motion owns the machines that imitate it.

---