● Maxinsights pays everyday people to wear cameras while they cook, clean, and do their jobs, then turns the footage into training data robots learn from

● MaxLoop, the company's proprietary AI platform, automates quality control of footage as it arrives and cuts the human review team by 90%

● MaxGlove, a sensor glove, records how hard a hand grips and how heavy an object is, which a camera can't see

SANTA CLARA, Calif.: Please replace the release dated Oct. 22, 2026 with the following corrected version due to multiple revisions.

The updated release reads:

WAYMO, GOOGLE AND MICROSOFT ALUMS RAISE $30M TO BUILD THE DATA LAYER FOR PHYSICAL AI

● Maxinsights pays everyday people to wear cameras while they cook, clean, and do their jobs, then turns the footage into training data robots learn from

● MaxLoop, the company's proprietary AI platform, automates quality control of footage as it arrives and cuts the human review team by 90%

● MaxGlove, a sensor glove, records how hard a hand grips and how heavy an object is, which a camera can't see

Maxinsights, the company building the data layer for physical AI, today announced $30 million in seed funding led by PeakXV, with participation from Tola Capital, South Park Commons, INCE Capital, and Alumni Ventures. The company also launched MaxLoop, an agent-in-the-loop system that scores and corrects data collection while it happens, and MaxGlove, a sensor glove for capturing tactile touch and force for a new era of data collection.

Investors put $47.4 billion into physical AI in the first half of 2026, nearly four times the total for the second half of 2025. The capital is flowing faster than the data that exists to train it. Robots can only learn to fold a shirt or load a dishwasher by watching people do it thousands of times, and nobody has captured that kind of work at anything close to the scale the models require.

The full data stack in one place

Competitors stage tasks in mock environments and ask operators to slow their hands so the footage is easier to process. Maxinsights records real work at natural speed in real homes, restaurants, and hotels, so models keep seeing new settings and unplanned moments. Every step from capture to delivery happens in-house

Maxinsights operates in five steps:

• Collect. First-person (egocentric) footage of real people working, captured in the wild by our own global network across varied places, objects, and tasks.
• Check. Automated and human review of every clip: capture quality, task completion, visibility, sync, metadata.
• Annotate. Proprietary models label each clip across four layers (environment, task, subtask, instruction) plus hand pose, body motion, and object interactions.
• Curate. Search and rebalance the corpus by task, scene, geography, behavior, complexity, or whatever a client needs.
• Deliver. Ship raw or processed datasets, then feed client and model feedback into the next collection cycle.

The company has captured more than 2 million hours, with 1.5 million hours model-ready and available for immediate delivery. Maxinsights serves over 30 customers, including frontier AI labs and foundation model companies, with eight on recurring contracts.

Two of the most closely watched robotics foundation models of 2026 were trained in part on Maxinsights data.

• Genesis AI: Maxinsights was the largest contributor of egocentric data used to train GENE-26.5, providing 200,000 hours of data.
• Dyna Robotics: Maxinsights is the largest data provider behind Dyna-2, Dyna's World-Action Model.

The captured footage comes from people getting paid to record work they were already doing. More than 5,000 contractors in six countries wear Maxinsights’ proprietary camera hardware while they load the dishwasher, make a smoothie, or wash a car. This captures clear footage compared to shaky iPhone or GoPro videos. Maxinsights hires them through staffing firms in each market and manages them directly rather than handing collection off to a vendor.

“We built the whole pipeline ourselves, from the hand-tracking models to the platform that scores every operator's footage as it arrives, because nobody was going to sell us the pieces that make messy real-world data usable,” said Yuguang (YG) Yuan, founder and CEO of Maxinsights. “Robots get deployed into real kitchens, not staged ones, and something unexpected happens every few minutes. Edge cases are the job, and the only way to get them is to record real work in real homes. Collecting one great hour of data is easy, and holding quality across millions of hours and thousands of people is what we built the company to do.”

Introducing MaxLoop and MaxGlove

Two products came out of the decision to build rather than buy. MaxLoop is the software that scores footage the moment it lands. MaxGlove is the hardware that captures what a camera can't see. Owning both means Maxinsights can change what gets collected and how it gets checked in the same cycle.

• MaxLoop. Real-world footage breaks most annotation tools. Hands disappear behind objects, several people's hands enter the same frame, operators wear gloves and rings, motion blurs. So Maxinsights trained its own hand-tracking models on footage it collected itself, and built MaxLoop to score every clip as it arrives. Operators now get feedback the same day instead of in one to two weeks, review costs are down roughly 80%, and the company processes 450,000 hours a month. The same models also pull more out of each clip: pose, object states, contact, and force. An hour of video that used to teach a model one thing can now teach it several.
• MaxGlove, which is being tested by Genesis AI, captures touch and force: how hard a hand grips, how much an object weighs, how contact shifts through a task. Foundation models need force data because vision alone can’t tell a robot how an interaction feels. Force and torque signals reveal contact, pressure, resistance, and slip, helping robots know whether they’ve grasped an object securely, how much force to apply, and when to adjust. No company has yet demonstrated a truly scalable way to collect force data: existing setups require lengthy or repeated calibration, while our device integrates with egocentric systems and can be deployed reliably across a wide range of environments.

Maxinsights manufactures its own because gloves on the market need long calibration cycles and yield only 10 to 20 minutes of usable capture a day. MaxGlove goes across the existing collection network, targeting 1 million hours of force data in 2027.

“Almost every company in this category is selling hours. Maxinsights is the only one that treated quality control as an engineering problem rather than a staffing cost, which is what lets them operate in real life, said Rohit Agarwal, Managing Director of Peak XV (ex-Sequoia India). “That decision is hard to copy after the fact, and it's why the labs that tried other vendors keep coming back to them.”

Yuan previously co-founded Mega Gear Robotics, an industrial robotics company that was acquired. Before that, as the youngest executive team leader at Eaton, he ran product groups with about $150 million in annual business and more than 1,500 employees. Kun Wang, co-founder and CTO of Maxinsights, was an early engineer at Google X/Waymo, then co-founded QCraft, which grew into a unicorn with more than 1,000 employees and deployed autonomous-driving technology in over 1 million vehicles. The broader team includes engineers from Google, Microsoft, Meta, Waymo, and Hulu.

About Maxinsights:

Maxinsights builds the data layer for physical AI, turning real-world human activity into robot-ready training data from capture to delivery. Rather than staging tasks in mock environments, the company records natural work in homes, restaurants, and hotels across eight countries, then runs quality control, annotation, and curation on proprietary models built for messy real-world footage. Maxinsights is headquartered in Santa Clara with offices in China, Indonesia and coming soon in Turkey and Mexico. Learn more at maxinsights.ai.

Fonte: Business Wire