Your workshop videos are worth a fortune for AI robotics

Large AI models can write, summarize, and code because they have ingested the text from the web. However, they remain very limited in a specific domain: the gestures of the physical world — manipulating an object, assembling a part, repairing a machine, picking a fruit. This data is not found online: it is produced in workshops, fields, warehouses, and kitchens, and therefore must be captured at the source.

The scale of the shortage is striking. According to an estimate by investor Bessemer (BVP), there are only about 300,000 hours of robotic manipulation data globally, compared to nearly a billion hours of internet video and tens of thousands of billions of words of text. This is an order of magnitude — a venture capitalist's estimate, not an audited figure — but it conveys the essential point: physical gesture data is rare, and therefore potentially valuable.

The companies concerned are those that operate in the real world: workshops and manufacturing, maintenance and repair, fleets equipped with onboard cameras, agriculture and agri-food, logistics and handling. What has value is the first-person gesture — egocentric video, filmed by a headcam or GoPro, showing hands at work, the grip, the sequence, and especially the difficult cases: defects, unforeseen events, corrections.

The market is structuring itself: robot manufacturers and annotators are seeking these corpora, and large-scale egocentric datasets are already being compiled. However, pay attention to the framework: filming a professional gesture raises questions about faces, voices, and locations. A proper approach — anonymization, consent, clear licensing — is what transforms these images into a secure asset. The first concrete step: verify if your activity already produces this data, through a free diagnostic on d-nvest.