Nomura analysts say data collection will be a critical constraint for
humanoid-robot training by 2026. They estimate about 10 million hours of data
are required per 100,000 units shipped. Nomura outlines four acquisitions
approaches, including non-physical data from wearables that capture human motion
and teleoperation of physical robots with synchronized camera feeds and
joint-state recording to train foundation models. The firm says an integrated
hardware-software loop covering collection, transmission, evaluation, training,
deployment and debugging would be a pronounced structural advantage for pure
data suppliers.
