Physical AI's missing training layer: humans, or remote robots?

Physical AI is not waiting on smarter models. It is waiting on a million hours of training footage, and that footage is a human problem. Today's embodied-AI training sources are thin: teleoperation requires owning a robot, hand-labeled video scales against labor cost, simulation runs into the realism gap. Each approach hits a different wall, and the field's open question is which side of capture gives first.

Ropedia, which raised $22M in pre-Series A funding this week, is one specific bet that the missing layer is first-person 360° video of people, captured by a wearable head-mounted camera called HOMIE. Co-founder and CEO Zhaoxi Chen frames the operational gap directly: "Delivering 1,000 hours of data is a totally different thing compared with delivering 1 million hours of data, because you need to make sure that everything is industry level of standards." The Ropedia thesis is not that the company has a million-hour dataset. It is that Ropedia is building the standard the million hours would have to meet.

The strong counter-reading: dedicated teleoperation fleets could reach million-hour scale on their own, leaving wearables as a niche layer. The evidence on the table is a young company, founded in 2025, with $30M total raised, on a thesis that the human eye is the missing sensor. Either way, the next phase of physical AI may be decided on the capture side — but this is Ropedia's inferred structural argument from stated limitations and company positioning, not an industry consensus.