hardware, training, fleet
Autonomy teams are often making the same wager with different chips.
Under the slogans, the sensor debate is a sequencing bet: which problem do you treat as truly hard, and which do you treat as merely expensive?
One camp says geometry should be measured directly so the team can spend its effort on behavior. The other camp says geometry from cameras is tractable, and the scarce work is semantic understanding — which has to be done either way, so do that work first and use cheaper sensors to buy scale.
The other camp says geometry from cameras is tractable, and the scarce work is semantic understanding — which has to be done either way, …
These are not opposite claims about physics. They are different judgments about order under limited time and money. Both camps agree that understanding scenes matters. They disagree about whether precise measurement is a shortcut worth its cost or a detour that delays the harder work.
The wager is also about data. Measurement-heavy designs can produce richer geometry per mile and fewer miles. Vision-heavy designs can produce weaker geometry per mile and many more miles. If the remaining problem yields to careful engineering on high-quality clips, the first bet looks better. If it yields to exposure to a huge variety of real streets, the second does.
The answer is not knowable from a comment thread. Both approaches already work in limited settings. The interesting test is which approach transfers to places it was not specifically prepared for — and that takes a long time to observe.