MaxInsights Scales Physical AI Dataset to 2M Egocentric Hours, Targeting 10M

Robot data company MaxInsights announced its Physical AI dataset has grown to 2M hours of egocentric human-experience video, with monthly output of 450K hours, calling it only the beginning with a next target of 10M cumulative hours. This follows Dyna-2's validation that 1M+ hours of first-person human video can scale robot learning—performance kept improving as data grew from 1K to 1M hours, with pre-training using no robot data at all—signaling that the robot data bottleneck is being dissolved by scaled real-world experience data.

Confirmed

• MaxInsights has accumulated 2M hours of egocentric data, produces 450K hours of video per month, positions itself as the data foundation for Physical AI, and targets 10M hours next.
• The company published the research report "Beyond 1M Hours: The Data Scaling Laws of Physical AI," systematically arguing for the next frontier of robot data.
• The report stresses that duration is the least important release metric: two recordings of equal length can differ in learning value by an order of magnitude—an hour of repeated pick-and-place under fixed conditions is far less valuable than an hour rich in bimanual manipulation and diverse contacts.
• Its expansion framework advances along two dimensions (scale and content quality), with the core formula: effective experience ≈ hours × information density per hour.
• Egocentric human video has become a key pre-training asset for robot manipulation skills, but works significantly better only after human experience is "translated" to the robot's own embodiment and movement patterns before training.

Why it matters

• Robotics has lacked a unified, scalable data substrate like text or images; MaxInsights' scaling-law framework gives the industry a methodology for shifting from stacking hours to competing on information density.
• As @rohanpaulai notes, beyond the million-hour mark the data race will shift from scale to per-hour experience quality and human-to-robot transfer efficiency, directly shaping pre-training paradigms for embodied intelligence models.

2026-09-22 ~ 2026-09-23 · 5 related posts

• Human video emerges as robotics pretraining substrate: effective experience = hours × information per hour — rohanpaul_ai · 2026-09-22
• Maxinsights report: Physical AI data scaling laws — after 1M hours, it's experience density — rohanpaul_ai · 2026-09-22
• Robotics data firm Maxinsights claims 2M hours of egocentric human experience — ChrisGPT · 2026-09-22
• [source] MaxInsights hits 2 million hours of egocentric video to fuel Physical AI scaling — CyberRobooo · 2026-09-23
• [source] MaxInsights pushes physical AI data to 2M hours of real-world experience, 450K hours produced monthly — CyberRobooo · 2026-09-23