MaxInsights Founder Yuan Yuguang on Producing First-Person Training Data for Robots

Chen Qian introduces MaxInsights, a company that supplies training data rather than building robots. Founder Yuan Yuguang says its annual recurring revenue grew from about $1 million last year to about $250 million, and that it has accumulated nearly two million hours of first-person human data through a global network of nearly 4,000 collectors.

Yuan says MaxInsights converts records of human activity into training data for leading model and robotics companies. He attributes its rapid growth to changing industry demand: customers have used large quantities of high-quality data to train models and observed new capabilities and performance.

Chen asks whether the industry previously lacked teleoperation data or first-person footage. Yuan says the company explored teleoperation, simulation, first-person video and gloves with customers. Teleoperation has a production ceiling, while simulation faces a gap between virtual and real environments.

The first order was for 200 hours of first-person data from a leading US model company, Yuan recalls. They settled on that direction because it was useful for training, could scale quickly and might turn tacit human skills into knowledge and capabilities for robots.

He uses folding clothes as an example: people fold garments in many ways, and training on a diversity of human methods might help robots handle different clothes more reliably and efficiently. Chen describes this as better generalization through data; Yuan says they saw early signs of that last year.

According to Yuan, the company went from an initial 200 hours of first-person data to about 20,000 hours accumulated last year and around 200,000 hours early this year. As volumes increased, the problem shifted from a simple shortage of data toward a shortage of high-quality data.

Collectors wear head-mounted devices to record everyday activity; the raw footage is then processed into training data. Chen notices six cameras on the previous device. Yuan says its roughly 270-degree field of view helps keep the hands in frame despite head and hand movement.

The device is intended to capture hands, body parts and the surrounding environment to support hand and pose tracking. Yuan says the company favors relatively professional collectors and teams, with direct management of quality, output and diversity.

Yuan compares the network to DoorDash or Meituan: it must organize many collectors while controlling quality and output. He argues that direct management helps meet customers’ requirements for scale, quality and diversity.

The newer device is lighter, lasts longer on a battery and calibrates more quickly. It reduces the camera count from six to four to optimize cost while retaining the necessary information. It is worn like a cap and connected through a phone app.

On capacity, Yuan says the company can collect up to roughly 450,000 hours a month and has nearly two million hours on hand. He describes it as a high-quality first-person dataset, prompting Chen to ask how customers can tell whether large amounts of footage are actually useful for training.

Yuan says quality includes accurate tracking, diverse data, sufficient volume, language annotations and meaningful tasks. He says tracking must reach millimeter-level accuracy; aimlessly using a phone or scratching oneself should not count as useful task output.

Turning raw video into a training-ready dataset requires a complex pipeline: hand and language annotation, video understanding and indexing, and quality checks. Humans also help validate the output, and their feedback can improve the models and processing pipeline.

Yuan says the company assigns confidence levels to data for different customer requirements but does not directly deliver low-quality data. Instead, it returns inadequate data to the pipeline to diagnose annotation and processing problems. He says the pipeline has gone through roughly six or seven iterations.

Chen asks how first-person data differs from UMI data. Yuan says a person collecting UMI data uses a gripper that can be mapped directly to a robot gripper, while bare-hand first-person data records natural hand movements. UMI can be used for mid-training and even post-training.

Gripper data suits robots with grippers as end effectors, he says, while bare-hand data can map to dexterous hands. Which type a customer uses, or how they combine them, depends on the robot and the training task.

Yuan says completing tasks with a UMI gripper is usually slower than using bare hands, but downstream processing is simpler. Bare-hand movements are more natural and less controlled: tracking becomes harder when hands are occluded, crossed or interacting with another person’s hands.

Customers need teleoperation, bare-hand first-person and UMI data, but Yuan observes the largest current demand for human hands. He says teleoperation is generally more expensive and first-person data sits in the middle; across the industry he estimates the latter may cost about half as much, depending on requirements.

Yuan argues that customers can generate substantial simulation data themselves, so selling simulation data alone is not his company’s main focus; simulation firms can still provide environments and assets. World-model and video-generation companies need real data, he says, and some are customers.

Chen asks whether video-generation companies that buy real data and sell generated data to robot makers could become competitors. Yuan sees a complementary relationship: one side supplies real data, the other generates more video, and the data company can annotate and process the results.

Yuan regards the company’s pipeline and proprietary ground-truth data as a moat. He says its models draw on years of accumulated data, and that hardware, collection, workforce management, processing models and delivery must form one vertically integrated chain. His belief that newcomers will struggle to catch up is his assessment of the competitive position.

Asked who collected the nearly two million hours, Yuan says the company directly manages nearly 4,000 collectors worldwide. He likens the network to a factory spread across many locations: it can send collectors new assignments for settings such as hospitals or care facilities and get feedback quickly.

Yuan says the team tried outsourced collection networks but found it hard to meet quality and output needs. In January 2026 it requested 100,000 hours from a provider, with a minimum of 20,000; the provider delivered about 18,000, prompting a change in operations.

Afterward the team organized resources around its devices: they must be accurate, correctly worn and well utilized. Chen summarizes the requirements as good hardware, training and management; Yuan adds that a platform must coordinate both people and equipment.

Yuan says fewer than 10 full-time collection operations staff use a platform with AI agents to help manage around 4,000 collectors. It can give feedback on quality and output, for example by alerting someone that a device is worn incorrectly. He calls the agents assistants.

The collection network spans the United States, Mexico, Indonesia, Vietnam, Thailand, China, Turkey and South Africa. Yuan says differences in food, clothing and living environments across countries help diversify the data.

Customers fall into two groups: major model companies and robotics intelligence companies. Yuan does not name specific customers; he says the company provides training data and wants to combine it with evaluation and deployment support for both groups.

Chen asks whether the same dataset can be licensed to multiple companies, like SaaS software. Yuan says the company licenses dataset use, the dataset grows daily, and a validated dataset can be reused by different customers.

How can models differ when customers use the same data? Yuan points to different model architectures, training recipes, adaptation methods, mid-training and post-training. In deployment, moving from 99% performance to a more reliable level requires further training and feedback.

He expects demand for data to grow, with some customers requiring post-training, tests and evaluation in specific environments. Chen notes that the industry is exploring a move from one million to ten million hours to test whether scaling effects continue.

Yuan estimates that producing ten million hours could require three to five times as many collectors, roughly 15,000–20,000 people. He says the challenge is maintaining quality, diversity and a reliable pipeline while adapting to changing customer requirements.

Yuan says the company has more than 30 customers with different detailed requirements, making continual communication necessary. Chen asks whether more specialized robotics scenarios will become more expensive to collect, as expert annotation has for advanced language models.

Yuan believes robotics data demand is moving from simple mockups and repetitive actions like wiping tables toward more complex, specialized tasks. For deployment, data and testing must also target specific environments, such as care facilities in different regions, to improve task success.

He says expanding from hundreds of hours to millions or tens of millions requires redesigned infrastructure, including upload networks and annotation. The company uses its own annotation models to reduce the cost of large-model calls while improving quality.

Chen asks whether robotics data could be exhausted like text on the internet. Yuan says different data types have different limits: with a fixed model architecture, gains in success rate may level off as data grows. Further improvements and harder tasks may require more specialized data.

Yuan expects the collection network to expand from general workers toward experts who can help customers address harder deployment cases. He believes those who help customers deploy faster and recoup their costs will be better placed to win trust and orders.

What value can data firms provide if scaling gains slow? Yuan proposes evaluation and deployment: test where a model fails in real environments and whether it can recover, then turn failures and recoveries into data for improving it.

He envisions combining data production, global operations, evaluation and deployment in more complete support for customers such as retailers. He describes the necessary capabilities as high-quality data processing, operations and continuous collaboration with customers.

Chen turns to Yuan’s entrepreneurial background. After graduation he considered consulting and automation, ultimately joining Eaton because he wanted hands-on work. The job took him to the US, Mexico and Japan and gave him global business experience.

Yuan worked in Japan for about four years and encountered the Toyota production system. He says quality can suffer when output increases, so teams must control and adjust before scaling again. He now applies similar quality-management ideas to a worldwide data collection network.

He lists operational problems: meaningless tasks, incorrectly worn equipment, occluded hands and repetitive settings can reduce useful data quality. Four people repeatedly folding the same garment in different corners of one room may boost apparent throughput without helping model generalization.

Around 2014 Yuan left Eaton to start a company. He compares the choice to a bird leaving a cage: freedom opens up a larger world, but food is not guaranteed. His first company tried to make lower-cost industrial robots to compete with more expensive incumbents.

His first robot lacked a finished casing and he considered scrapping it. A university teacher then showed interest in a robot that could be programmed at a low level, revealing a use for the open platform in universities and vocational schools.

He says the robot sold for nearly RMB 200,000 per unit, still less than some international brands at comparable payload. The company later moved from hardware to computer vision, encountering the difficulty of collecting reliable images for precise manipulation of products with many shapes.

He eventually sold the company. Looking back, he advises founders to consider their business model and path to profitability early, rather than leave the cage first and search for direction while resources run short.

The company pivoted several times, including to connected factory monitoring: it rapidly connected hundreds of robots, but customers would not pay. Yuan says a stated need is not the same as willingness to buy; a real pain point can drive payment even before a product is polished.

He reflects that he asked only a few acquaintances if they wanted to try the service and initially offered it for free. He later found limited willingness to pay, and the people he asked were not necessarily decision-makers. Explaining the price upfront might have produced a different response.

Before founding his next company, he studied for an MBA at UC Berkeley. After seeing a robot fold clothes in 2023, he sensed change and sent cold emails to people in robotics, eventually consulting nearly 200 people and raising his response rate from around 10% to 40%.

Yuan says close exchange between Bay Area academia and industry helped shape his early view of robotics and the data bottleneck. He describes the company’s direction as a combination of research and engineering: advance frontier technology while scaling data production.

Asked how he identified data as a real pain point, Yuan says he looked at data needs in autonomous driving and digital AI and spoke directly with people making frontier research decisions. He also stresses that the company’s ultimate opportunity depends on broad deployment across embodied AI.

Asked whether selling picks and shovels is less compelling than building robots, Yuan says a team should start from its own strengths. He sees his contribution as combining engineering, research and operations to help customers scale, rather than personally doing the deepest AI research.

Chen asks about competition from data companies in lower-cost regions. Yuan says operations and cloud costs make up much of the cost per hour; the company has offices in Silicon Valley, Suzhou and Indonesia focused respectively on frontier technology, tool development and operations management.

He says proximity to US customers helps the American team provide rapid service and iteration, and the company hopes eventually to open a European office. Closer feedback can help evaluation and deployment; serving global customers requires global operations.

On the cold start, Yuan says the team believed model architecture, cloud and data were central to the industry, but had to keep reassessing specific routes such as simulation and teleoperation. He believed customers could generate simulated data themselves, making data sales alone difficult.

The company built a teleoperation setup in a cofounder’s garage and took orders, but encountered equipment reliability and maintenance problems. Simple pick-and-place tasks required repeated resets that could take as long as the action itself, limiting useful output and pushing the team toward more scalable options.

He expects the company to need evaluation and deployment capabilities alongside more data, helping customers with pre-training, mid-training and post-training in real settings. The immediate challenge is to improve quality at scale while keeping operations and capacity flexible.

Chen raises reports of companies buying data from subcontractors and reselling it in a new package. Yuan opposes the practice and says MaxInsights collects and owns its own data. He argues that controlling provenance, the pipeline and standards is essential to consistent delivery.

Over the next year Yuan expects robotics data demand and deployment in specialized settings to grow. He says the company wants to become a data partner rather than just a supplier, iterating with customers; prices must ultimately reflect value delivered through model improvements.

He argues that cheap, poor-quality data may increase compute costs or even degrade a model and require retraining. The company is considering tiered pricing for different forms of data, including hands, upper body, whole body and glove-based collection.

Yuan shows a sensor-equipped glove and discusses collecting tactile or force-related information. He says reliable collection at scale remains difficult for new modalities, and different data types may command different prices.

Chen asks whether robot deployment will take longer than expected, as autonomous driving has, because of safety and edge cases. Yuan sees a difference: a robot can first solve a specific task in a limited setting, such as folding clothes continuously in a hotel, before expanding to other tasks.

Yuan expects continued rapid iteration among world models, vision-language-action models and world-action models. This is his optimistic industry forecast; it does not establish that general-purpose capability or deployment at scale has already been proven.

He argues that training models will continue to require data, while data companies must understand changing needs and combine collection, processing and service. Working with multiple customers, he hopes to form reusable data products and industry standards.

Yuan closes by tracing industry data needs from hundreds to tens of thousands and hundreds of thousands of hours. He welcomes people from autonomous driving, traditional robotics and control, and expects repeated iterations before robots of any form provide value across a wide range of settings.

Chen concludes that million-hour first-person datasets are an important direction for robotics training, but scaling effects still need testing at larger sizes. She identifies tactile and other modalities, evaluation, post-training and real-world deployment as areas to watch.