Maxinsights Raises $30M Seed for Physical AI Training Data

Client feedback feeds the next collection cycle at Maxinsights, giving a robotics lab influence over what somebody records tomorrow. The Santa Clara, California company has built its physical-AI data business around that connection between the person doing a task and the researcher trying to learn from it.

Maxinsights announced a $30M Seed round on October 22, 2026, led by Peak XV, with Tola Capital, South Park Commons, INCE Capital and Alumni Ventures participating. Founder and CEO Yuguang Yuan and co-founder and CTO Kun Wang are developing a pipeline that carries human activity through capture, checking, annotation, curation and delivery.

The company also introduced MaxLoop, its incoming-data checking system, and MaxGlove, a force-sensing glove being developed with Genesis AI. Together, those products put two consequential decisions inside the supplier's business: which recordings deserve to reach a training run, and which physical information the recording should contain. For robotics buyers, the financing draws attention to the operations required before a dataset becomes useful.

What Maxinsights is building with its $30M Seed

Maxinsights records people carrying out work in real environments, then prepares the material for model developers. A dataset can be curated around a customer's task or setting, with feedback informing subsequent collection. That continuing relationship gives the supplier a role in deciding how the next batch should differ from the last.

Consider an illustrative request for more examples of handling an unfamiliar object. The researcher needs to specify the useful interactions, the collection team needs to capture them, and the review process needs to identify whether the required evidence survived. Paying for another recording session only helps when those people agree on what the session is supposed to produce.

That is an operating assignment with several opportunities for an expensive misunderstanding. The person performing the task can complete it successfully while producing a poor view of the movement. The processing system can generate a clean-looking output that still lacks the observation a particular research question requires. A supplier accepting responsibility across those stages has more ways to diagnose the problem, alongside more work to coordinate.

Maxinsights reports more than 2M recorded hours and describes 1.5M as model-ready. Those company-reported figures describe different stages of inventory. The announcement leaves revenue, valuation and the cost of producing a usable hour undisclosed, so the financing provides no basis for calculating the economics of that preparation.

Why checking the recording changes the collection job

MaxLoop is designed to assess footage as it arrives and return feedback to the collection process. In the company's account, the system addresses difficult recordings in which hands become obscured or movement is hard to interpret. Finding a problem promptly can make the next instruction to a collector more useful.

For a lab purchasing data, that timing creates a practical distinction. Discovering an inadequate camera view after a collection campaign has ended can leave a decision about whether to repeat the work, change the training specification or accept a gap. Detecting it during collection gives the supplier a chance to adjust while the people and equipment are still available.

The commercial opportunity sits in reducing avoidable work across that relationship. A buyer can evaluate whether feedback is specific enough to change collection, how questionable material is handled and how the delivered dataset reflects the original request. These are purchasing considerations implied by the pipeline, rather than published performance guarantees for MaxLoop.

Yuguang Yuan and Kun Wang have therefore put engineering inside an activity that can look like a staffing service from a distance. The collection team produces observations; the software helps determine which observations belong in the delivery. As the customer changes the research task, both sides of that operation have to respond. An extra hour has value only in relation to the learning job for which it was collected.

Human-video research gives the buyer a sharper specification

Dyna Robotics' Dyna-2 technical report offers primary research context for the interest in human recordings. The August 2026 report describes pretraining on more than 1M hours of egocentric human video and reports improvements in transfer to robot data as the human-data scale increases. It also finds that the learning objective matters to that transfer.

Those results give a data purchaser reason to ask how a proposed collection fits the model being trained. They do not establish that any large collection will deliver the same result. The research links the data choice with the training approach, keeping the buyer's technical specification central to the purchase.

Maxinsights' own September research discussion considers both the amount of experience and the physical information retained within it. For a commercial supplier, that introduces a planning question before the cameras go out: which additional observation would be worth paying to obtain? Repeating familiar activity and capturing a missing interaction can consume similar recording time while serving different research needs.

That perspective also explains why customer feedback belongs close to collection. A model developer may learn that the next dataset needs a different view, a different environment or an additional measurement. A supplier able to translate that request into a collection plan participates in the customer's research process, with the quality of that translation becoming part of the service being purchased.

Force sensing adds another customer decision

MaxGlove extends the collection proposal into touch and force. Maxinsights says it is developing the device with Genesis AI and is targeting 1M hours of force data in 2027. That number is a future collection objective, separate from the video inventory already reported.

For the buyer, the new modality brings questions about which interactions require direct measurement and how those signals align with the rest of the recording. For the collection operation, it adds equipment that must remain useful during ordinary work. A richer recording specification can create more valuable evidence while also increasing the preparation required to produce it.

The relationship between physical measurement and learning is appearing elsewhere in the robotics market. DevCuration's coverage of Enact Intelligence's force-aware robotics work follows a company developing the machine and its learning system together. Formic's factory-automation expansion examines the service responsibilities that follow equipment into production. Maxinsights occupies the data-supply part of that wider operating chain.

Its next delivery will carry the consequences of choices made before recording began: what the lab asked to learn, what the equipment could observe and what the collection team was instructed to preserve. As force sensing joins that process, the discussion between the researcher ordering the data and the person preparing the recording becomes more specific about the physical work still missing from the model's education.