Maxinsights raises $30 million to expand data infrastructure for Physical AI

Robotics and Automation News reported on October 28, 2026, that Maxinsights had raised $30 million in seed funding to expand its data infrastructure for Physical AI. The company says it has collected more than 2 million hours of video documenting everyday tasks, of which more than 1.5 million hours are ready for model use.

The company reportedly announced a setup that combines video of real-world activities with MaxGlove, which records grip strength and object weight, and MaxLoop, which handles video quality. The design supplements information that cannot be obtained from video alone with other sensors, indicating that businesses involved in collecting and processing data for robot training need to assess data types and quality control together.

Alongside the $30 million raised, the key aspect of the company's business is that it puts data collection, quality control and tactile-data capture into a single workflow. According to the report, Maxinsights has collected more than 2 million hours of activity video, of which more than 1.5 million hours are ready for model use. These figures, however, were reported as the company's own claims; the article does not establish the criteria for deeming data ready for model use or provide an independent evaluation. The allocation of the funding has also not been disclosed, so it is not possible to determine how the $30 million will be divided between expanding data collection and product development. In terms of how the data is produced, the report describes a setup that combines video documenting everyday tasks with quality control by MaxLoop and adds information on grip strength and object weight captured by MaxGlove. Recording physical quantities that are difficult to capture from video alone with a separate sensor shows that video data by itself may not provide all the information needed to train robot actions. Further evaluation will need to examine how sensor data is matched with video, whether tasks or environments are underrepresented, and how quality assessments affect actual training. In an article we published the previous day, Google DeepMind unveils virtual-world generation AI “Genie 4,” with support for audio and conversation, we noted that generating virtual worlds in response to user actions is intended for uses such as robot training, while changes in objects' appearance and position during long generation sessions were cited as a challenge. Generating virtual environments and collecting data by recording people's work in the real world are different methods, and it cannot be said that either replaces the other. What can be gleaned from this report is that the company is seeking to include tactile information, in addition to video of real-world tasks, in the data it collects. The potential industry implication is that the process of selecting and processing data into a form usable by models could be positioned as a standalone infrastructure, rather than data collection alone. However, the total of more than 2 million hours does not reveal the range of tasks, recording conditions or reusability of the data. What should be examined next are the criteria for deeming data ready for model use, data rights, the synchronization accuracy between MaxGlove recordings and video, and evaluations of robot actions after training. The report also does not disclose the specific uses of the funding or what kinds of model training have used the recorded data.