Maxinsights Raises $30M After Collecting 2 Million Hours of Real-World Robot Training Data

Santa Clara-based Maxinsights has raised $30 million in seed funding to build a data infrastructure platform for physical AI, turning recordings of everyday human activity into datasets used to train robotics foundation models.

The round was led by Peak XV, with participation from Tola Capital, South Park Commons, INCE Capital and Alumni Ventures.

The company reports collecting more than two million hours of real-world footage, including approximately 1.5 million hours prepared for model training.

Maxinsights is betting that robots will need access to large volumes of real physical experiences, even as foundation models become more capable.

Large language models learned from vast quantities of text already available online.

Robotics developers face a different challenge.

The internet contains relatively little structured information showing how people physically manipulate objects from a first-person perspective.

A conventional video might show someone cooking, but it rarely captures every detail needed to understand hand positions, contact forces and object movement.

Robots must learn these relationships to perform practical tasks.

Maxinsights collects recordings of people carrying out normal activities in homes, restaurants, hotels and other environments.

Examples include preparing food, cleaning surfaces, handling objects and performing repetitive operational tasks.

The company then processes those recordings into structured training data.

Maxinsights operates a distributed network of contractors using proprietary camera equipment.

Workers wear cameras while performing tasks they would otherwise complete during their daily routines.

The equipment captures first-person video of physical interactions rather than relying exclusively on staged demonstrations inside robotics laboratories.

According to the company's funding announcement, more than 5,000 contractors were participating across six countries.

Its website separately describes an eight-country collection network. The figures reflect different reported measures of active workers and geographic operations rather than a verified count of identical deployments.

The company must coordinate equipment, worker training, recording procedures, quality standards and data processing.

That makes the business partly a software platform and partly a distributed physical operation.

One of its new products is MaxLoop, an agent-in-the-loop system designed to inspect training footage shortly after it is recorded.

Traditional data collection pipelines often rely on human reviewers to identify unusable recordings.

Problems such as incorrect camera positioning, incomplete tasks or poor visibility may only become apparent days later.

MaxLoop automates parts of that review process.

The software analyzes incoming footage, identifies potential quality problems and provides feedback while workers can still repeat the activity.

According to Maxinsights, the system has reduced its human review requirements by approximately 90%.

The company also reports shorter feedback cycles, reducing delays that previously could extend beyond a week.

These are internal operational metrics and have not been independently benchmarked across competing data collection platforms.

Maxinsights has also introduced MaxGlove, a sensor-equipped glove developed in collaboration with Genesis AI.

The product addresses information that ordinary video cannot reliably capture.

A camera can show a hand lifting a cup, but it cannot directly measure how much force the fingers apply.

It also cannot fully describe contact pressure or how the grip changes when an object moves.

MaxGlove is intended to record tactile and force information alongside video.

That could help robotics models learn tasks involving delicate materials, variable object weight and complex manipulation.

The company aims to collect up to one million hours of force-related data during 2027.

This remains a target rather than a completed dataset.

Maxinsights says it serves more than 30 customers, including AI research laboratories and robotics foundation model developers.

Eight customers reportedly have recurring contracts.

The company identifies Genesis AI and Dyna Robotics among the developers using its datasets.

According to Maxinsights, it supplied approximately 200,000 hours of egocentric data for Genesis AI's GENE-26.5 model and was the largest data provider behind Dyna Robotics' Dyna-2.

The company has not disclosed revenue, gross margins or the value of individual customer contracts.

Maxinsights competes with specialized data collection providers and larger AI data businesses such as Scale AI and Surge AI.

It also faces a different form of competition: robotics laboratories collecting their own proprietary training data.

Some researchers are pursuing approaches that require fewer real-world demonstrations, using simulation and reinforcement learning to reduce data collection costs.

Machenta, for example, is developing a foundation model trained entirely in simulation before receiving a real-world task demonstration.

These approaches are not necessarily mutually exclusive.

Real-world data may remain important for validating model performance even if simulation reduces the amount needed for initial training.

The central question is whether its datasets improve the real-world capabilities of customers' robots.

Collecting millions of hours is an operational achievement, but data volume does not automatically translate into better models.

Quality, diversity, annotation accuracy and relevance to actual robotic tasks are equally important.

The company also faces privacy and consent obligations when recording people inside real environments.

Maxinsights' opportunity is to become a data infrastructure provider for physical AI. Its long-term advantage will depend on demonstrating that its datasets improve robot performance enough for customers to continue purchasing them.