Mecka Raises $60M for the Data Layer Behind Humanoid Robot Learning

The Mecka Story: Recording Work Before Automating It

On its current website, Mecka describes itself as a data, evaluation, and deployment layer. It explicitly says it does not build robots. Its offerings include an iOS capture app, tools for browsing and querying datasets, and enterprise evaluation and deployment support. Those are company descriptions of its products and role; they should not be mistaken for independently measured outcomes.

That role addresses a practical purchasing problem. A robot buyer needs more than a convincing demonstration. Someone must specify the task, capture relevant examples, check the training material, integrate the machine into a workflow, and measure whether it keeps working. A data supplier that also participates in deployment can, in principle, follow failures back to the conditions that caused them. That is our interpretation of the business model, not evidence that the complete loop already works at every customer.

Consider a simple packing station. A recording may show where a person placed an item, but the robot must still determine how to grasp it, how to avoid damaging it, and what to do when the box moves. More examples help only if they preserve information relevant to those decisions. Counting recording hours without examining their contents risks confusing a larger library with a better learning resource.

The relevance to humanoids is therefore specific. Human demonstrations may offer a way to gather examples of manipulation before an expensive robot is available at every collection site. They do not automatically teach balance, safe whole-body motion, or recovery from a failed grasp. Each capability needs evidence at the level where it will be used.

Under the Hood: Human Data Still Needs a Transfer Test

The July 7 version of the EgoVerse paper reports 1,362 hours, approximately 80,000 episodes, 1,965 tasks, and 2,087 demonstrators. Its experiments generally found benefits from human data, conditional on alignment with robot learning. These are research findings from particular tasks and embodiments, not a universal humanoid success rate.

Alignment means making the human example and robot objective comparable. If the demonstrator approaches a task in a way the robot cannot reproduce, imitation can become a misleading target. A buyer should ask which portions of a recording become observations, which become intended actions, and how the resulting policy is evaluated. A policy is the learned rule that turns the robot's observations into its next actions.

DROID's project provides a useful contrast: it gathered robot demonstrations using a shared Franka Panda setup across 13 institutions. Its researchers report improved manipulation performance and robustness from their dataset. DROID is older context, not an October announcement or a Mecka product. Task definitions and collection methods differ, so the table cannot tell a customer which dataset will work better for its machine.

Version control also matters. The EgoVerse repository's July change log documents mandatory camera calibration information and changes to how human embodiments are represented. This is a concrete reminder that usable data include conventions and processing history. Two folders with the same number of clips may not be interchangeable. Teams need to preserve the exact data version, training configuration, and evaluation conditions behind a result.

The live project site now displays 4,003 demonstration hours while retaining approximately 80,000 episodes in its snapshot. We use the fixed paper figures in the comparison because the site and July paper describe different snapshots. The displayed growth does not establish a proportionate gain in robot capability.

Mecka: The Full Picture Behind the Financing

Mecka names NVIDIA, Microsoft's M12, Qualcomm Ventures, and Samsung among new investors, with continued backing from Kindred, Framework Ventures, and Neo. It says it exceeded a $100 million revenue run rate in June and projects $300 million by year-end. Those revenue figures and the projection are company claims, not audited annual revenue or evidence of robots deployed at that scale.

The distinction changes the interpretation. A run rate annualizes a pace of business at a point in time. It does not disclose the revenue already recognized over a full year, the share attributable to data rather than integration, or the costs of supplying either. Investors' participation likewise establishes financial backing, not validation of every technical claim.

There is a more tangible organizational signal on the careers page checked October 12: 50 advertised roles across 11 teams. These include camera and sensor calibration, hand tracking, capture programs, reliability testing, and forward-deployed robotics engineering. Listings show intended hiring, not filled positions, but their mix helps explain why the company calls itself infrastructure rather than simply a video marketplace.

Those roles span the work required to turn collection into a service: hardware must record consistently, research tools must interpret the recordings, and customer teams must connect outputs to real operations. The commercial challenge is making that chain repeatable without requiring a bespoke engineering project for every site.

What This Means for Humanoid Buyers

The financing is a reason to watch a supplier, not a reason to shorten an acceptance test. Our assessment is that customers should judge a data program by changes in completed work: whether the robot succeeds more often, needs fewer interventions, and recovers safely when conditions change. Those measurements should use tasks and environments excluded from training.

An evaluation should count attempted tasks as well as completed ones. It should record resets, intervention time, damaged items, and whether a person selected an easier case after a failure. A tidy highlight reel can omit all of these. These are proposed buyer questions, not measurements reported for Mecka.

Data collection also creates responsibilities toward the people and places being recorded. A buyer should establish permission to capture and use the material, retention terms, access controls, and how unrelated people or confidential objects are handled. A technical dataset description does not, by itself, answer the commercial permissions question.

What's Coming Next

The next useful disclosure would connect a frozen training dataset to a named deployment and repeatable before-and-after evaluation. Readers should look for the robot platform, task boundaries, trial count, failure definition, and intervention policy together. A result without those details is difficult to compare with another supplier's claim.

For the data business, repeatability will be the harder test than collecting another impressive volume of footage. Can a new customer reproduce the benefit without a lengthy custom project? Can the supplier explain which examples changed behavior? Can updates improve performance without breaking previously reliable tasks? The public funding announcement does not answer these questions yet.

Frequently Asked Questions

The Next Test Is Useful Work

Mecka's financing gives the humanoid sector another well-funded effort aimed at the gap between demonstrations and deployment. The evidence warrants attention to the data supply chain, alongside the robot body and model. The decisive commercial result will be a documented improvement in real tasks that another team can reproduce.