Tesla FSD (Hardware 4, latest build) still can't handle airport pickups competently—a critical blocker for steering-wheel-less Cybercabs launching next year.
Tested at SFO. The nav system has zero semantic understanding of airport infrastructure:
• Can't differentiate arrivals vs departures, domestic vs international
• No gate-level routing ("Terminal 1" means nothing when Delta has 8 doors)
• Doesn't know inner/outer loop logic for circling
• Failed to locate cell phone lot correctly, tried routing into long-term parking
• No "pull over at door X" command
• Zero flight context integration despite wife having Tesla app with location data
Path planning is also broken—stays in right lane (gets blocked by stopped cars) instead of using left lane then merging right near target door like humans do.
Grok can't compensate. Asking "navigate to Delta arrivals" returns irrelevant POIs like Delta Sky Club (inside security).
This isn't edge case stuff. SFO sees thousands of Teslas daily. If a Cybercab (no manual override) gets stuck in wrong terminal loop or can't find pickup zone, passengers are screwed.
The irony: rest of 45min drive was flawless. FSD crushes highways and city streets but completely falls apart in structured environments that require contextual reasoning about human infrastructure.
Real AGI would infer "picking someone up = arrivals level, specific airline = specific door cluster, early arrival = cell lot first." Current system has zero world model for airports.
Data isn't the problem (tons of SFO trips logged). Architecture is—nav system needs semantic layer that understands facility types, not just lat/long waypoints. Until then, robotaxis will be unusable for one of the most common autonomous use cases.
Tested at SFO. The nav system has zero semantic understanding of airport infrastructure:
• Can't differentiate arrivals vs departures, domestic vs international
• No gate-level routing ("Terminal 1" means nothing when Delta has 8 doors)
• Doesn't know inner/outer loop logic for circling
• Failed to locate cell phone lot correctly, tried routing into long-term parking
• No "pull over at door X" command
• Zero flight context integration despite wife having Tesla app with location data
Path planning is also broken—stays in right lane (gets blocked by stopped cars) instead of using left lane then merging right near target door like humans do.
Grok can't compensate. Asking "navigate to Delta arrivals" returns irrelevant POIs like Delta Sky Club (inside security).
This isn't edge case stuff. SFO sees thousands of Teslas daily. If a Cybercab (no manual override) gets stuck in wrong terminal loop or can't find pickup zone, passengers are screwed.
The irony: rest of 45min drive was flawless. FSD crushes highways and city streets but completely falls apart in structured environments that require contextual reasoning about human infrastructure.
Real AGI would infer "picking someone up = arrivals level, specific airline = specific door cluster, early arrival = cell lot first." Current system has zero world model for airports.
Data isn't the problem (tons of SFO trips logged). Architecture is—nav system needs semantic layer that understands facility types, not just lat/long waypoints. Until then, robotaxis will be unusable for one of the most common autonomous use cases.



