Mecka AI Raises $60M to Train Robots on Human Motion

The hardest problem in robotics is not the robot. It is the data. A startup called Mecka AI just raised $60 million on a bet that the fastest way to teach machines how to move is to record how humans already do it, using nothing more exotic than body sensors and an iPhone strapped to a willing participant. The company is not building a better arm or a cheaper actuator. It is building the training corpus that every humanoid maker is desperate to own.

What Actually Happened

Mecka AI confirmed it has raised $60 million across two rounds, a $25 million Series A closed in November 2025 and a $35 million follow-on that landed this week. The round was led by Framework Ventures, a firm best known for crypto investing, with participation from Menlo Ventures, SV Angel, and Kindred Ventures. Angel backer Ted Xiao, a former Google DeepMind researcher and a founding member of Jeff Bezos's Project Prometheus, also wrote a check. The company declined to disclose a valuation, which usually signals a founder who expects the next mark to be much higher.

The product is conceptually simple and operationally hard. Mecka pays people to wear a rig of body sensors, pair it with an iPhone, and go about ordinary physical tasks: folding laundry, opening doors, walking across uneven ground, manipulating objects with their hands. The sensors capture joint angles, gait, and fine hand gestures at high fidelity. That motion stream becomes labeled training data for embodied AI models, the kind that sit inside a humanoid and decide how to bend a wrist or shift weight without falling over.

The company is run by chief executive Josh Gao alongside co-founders Mogen Cheng, Jason Chong, and Duy Nguyen. It employs 40 people and, according to Gao, is already projecting a $100 million annual run rate based on contracts it has signed, though the customers remain unnamed. For a company that has existed for barely a year and sells a raw input rather than a finished robot, a nine-figure projected run rate is the number that made investors lean forward.

Why This Matters More Than People Think

Every conversation about humanoid robots fixates on hardware, on the elegance of a hand or the torque of a leg. The real constraint is that robots have almost no experience. Large language models were trained on the entire public internet, trillions of tokens of human text sitting there for free. There is no equivalent corpus of physical motion. A robot cannot read its way to dexterity. Someone has to generate the demonstrations, frame by frame, joint by joint, and that generation is slow, expensive, and physically bounded by how fast a body can actually move.

Mecka's wager is that this data layer, not the chassis, is where durable value accrues. If the company can build the largest high-quality library of human physical motion, it becomes a supplier to every humanoid maker rather than a competitor to any one of them. That is the picks-and-shovels position in a gold rush, and it is historically the position that survives when the miners go bust. The $100 million projected run rate suggests robot companies are already willing to pay rather than build this capability in house.

There is a second-order reason this matters. The dominant method for collecting robot training data today is teleoperation, where a human wears a controller and puppeteers the robot through a task while sensors record the result. Teleoperation is accurate but brutally slow and tied to the specific robot being driven. Mecka decouples data collection from any particular machine. A person folding a towel generates motion that can, in principle, train any humanoid with roughly human proportions. If that transfer works, the economics of robot learning change overnight.

Consider the cost structure. A teleoperation session requires a physical robot, a trained operator, a controlled environment, and time measured in the same minutes the task itself takes. A Mecka contributor needs a sensor suit and a phone, can work in their own kitchen, and generates data wherever ordinary life already happens. The marginal cost of one more hour of human motion is a fraction of one more hour of teleoperated robot motion, and it scales horizontally across thousands of people at once rather than serially through a finite fleet of expensive machines. That cost asymmetry is the entire investment thesis compressed into a single line, and it is why a data-first company can outrun better-funded hardware rivals on the one axis that matters most.

The Competitive Landscape

Mecka is entering a field that is suddenly crowded with money. Generalist AI just raised $400 million at a $2 billion valuation to build physical agents, Skild AI tripled to a $14 billion valuation, Physical Intelligence doubled to an $11 billion bet, and Figure AI sits at a $39 billion valuation chasing a humanoid for the home and factory. Each of those companies needs exactly what Mecka sells. The competitive question is whether they buy the data or hoard it as proprietary advantage.

The closest analogues Mecka itself points to are Wayve, which collects real-world driving data to train autonomous vehicles, and MicroAGI, which records home-cleaning demonstrations. The deeper historical parallel is Scale AI, which turned the unglamorous work of data labeling into a multi-billion-dollar business by becoming the neutral supplier to every self-driving and computer-vision team at once. Scale proved that the company selling the training input can be worth more than many of the companies consuming it. Mecka is explicitly running the Scale playbook for physical motion.

The risk to that playbook is vertical integration. Tesla is capturing Optimus motion data from its own factories, 1X is doing the same, and Figure has the capital to build its own collection pipeline. If the largest humanoid makers decide motion data is too strategic to outsource, Mecka's addressable market shrinks to the long tail of smaller robotics firms. The company's defense is scale and neutrality: a shared corpus larger and more diverse than any single robot maker could economically build alone, sold to everyone on equal terms.

The Scale AI parallel cuts both ways, and the cautionary half is instructive. Scale built an empire as the neutral data supplier, then watched its largest customers, including OpenAI and Google, begin building in-house labeling pipelines once the work became core to their advantage. Neutrality is a powerful position right up until your customers decide the input is too important to rent. Mecka is betting that physical-motion collection stays operationally painful enough that even Figure and Tesla would rather pay than staff a global network of sensor-wearing contributors. That bet looks reasonable today and could look naive the moment a humanoid maker proves it can collect at scale internally.

Hidden Insight: The Embodiment Gap Nobody Wants to Price

Here is the uncomfortable question buried under the $60 million: does human motion data actually transfer to robots at all? A human shoulder has a different range of motion than a robotic one. Human tendons store and release energy in ways no current actuator replicates. When a person folds a towel, they exploit skin friction, micro-adjustments, and a lifetime of proprioceptive feedback that a robot simply does not have. Recording the human motion is the easy part. Bridging the gap between a human body and a mechanical one, the so-called embodiment gap, is the part that has humbled every robotics lab for two decades.

The bear case, however, is straightforward: critics argue that human-motion capture produces beautiful data that robots cannot use directly, and that the genuinely useful signal still has to come from the robot's own body through teleoperation or reinforcement learning in simulation. Skeptics point out that DeepMind, Tesla, and others have poured resources into motion retargeting with mixed results. If retargeting human demonstrations to robot morphology remains lossy, Mecka is selling a high-quality input to a process that still leaks most of its value on the floor.

Mecka's counter is that scale and diversity change the math. A small set of clean human demonstrations may not transfer well, but a massive, varied corpus gives learning algorithms enough coverage to infer the underlying task structure rather than the exact joint trajectory. The model learns what folding a towel means, not just how one specific person did it once. Whether that holds is the empirical question the next 18 months will answer, and it is the question every customer signing those contracts is implicitly betting on.

There is also a quieter strategic signal in who led this round. Framework Ventures is a crypto fund, not a robotics specialist. Its presence suggests that capital which chased decentralized infrastructure is now rotating into physical AI infrastructure, looking for the next network-effect business. A motion-data corpus that gets more valuable as more contributors and more consumers join it has the same flywheel logic that crypto investors spent years hunting. That framing, data as an appreciating network asset, may matter more to Mecka's eventual valuation than any single robotics benchmark.

The deepest insight is about timing rather than technology. Foundation models for language hit their inflection only after the training data became abundant and cheap, not when the architectures were first invented. Robotics may be sitting at the exact pre-inflection moment language was in around 2018: the algorithms exist, the compute exists, and the missing ingredient is a large, clean, diverse dataset of the right modality. If that analogy holds, the company that solves physical-motion data does not just sell an input, it sets the clock on when general-purpose robots actually work. Mecka is small, unproven, and one of several, but it is aimed squarely at the bottleneck the entire field privately admits is the real one.

One more structural point deserves emphasis before looking ahead. The data Mecka collects is not a static asset that depreciates the way a labeled image set does once models saturate it. Human motion is effectively infinite in its long tail: every new task, every new object, every new environment generates fresh demonstrations that have never been recorded. That means the corpus can keep compounding in value for years rather than hitting a ceiling, and it means a contributor network, once built, becomes a renewable supply line rather than a one-time scrape. Investors who lived through the exhaustion of the public-text corpus, where the open internet ran dry as a training source, are acutely aware that physical motion has no such hard limit. That scarcity-resistance is precisely what makes a motion-data business attractive in a year when everyone is asking where the next trillion tokens come from.

What to Watch Next

In the next 30 days, watch whether Mecka names a single marquee customer. The $100 million run rate is built on signed contracts, but anonymous contracts invite skepticism. A named humanoid maker on the customer list would convert the projection from a pitch-deck figure into a validated market. Also watch for any published transfer results: a credible demonstration that Mecka-trained motion improves a real robot's task success rate would silence the embodiment-gap critics faster than any funding announcement.

Over 90 to 180 days, the metric that matters is corpus size and diversity. How many hours of human motion has Mecka actually captured, across how many distinct tasks and body types? A motion-data business lives or dies on coverage, and the company that crosses the threshold where its data demonstrably generalizes will pull ahead decisively. Watch also for vertical-integration moves from Figure, Tesla, or 1X, which would signal that the biggest players have decided to build rather than buy, capping Mecka's ceiling.

The longer arc to track is whether a true data marketplace emerges for physical AI, the way Scale and others became neutral suppliers to computer vision. If Mecka, MicroAGI, and others standardize formats and sell motion the way cloud providers sell compute, robotics could get its own infrastructure layer within two years. If instead each humanoid maker walls off its own data, the field stays fragmented and slow. Mecka's $60 million is, in effect, a bet on the open-market outcome.

The breakthrough in robotics will not be a better hand. It will be the moment someone finally builds the training set the field has never had.

---