conceptautonomous-driving
Vision-Only vs Sensor Fusion
Notes
Vision-Only vs Sensor Fusion
One-line summary: The central architectural fork in autonomous-driving perception — cameras-only (Tesla) vs redundant LiDAR + radar + camera + HD maps (everyone else).
The insight
Major autonomous-driving programs have split into two camps on the perception stack. tesla-fsd remains the only major OEM pursuing vision-only; waymo, mercedes-drive-pilot, GM, BMW, and Lucid have all committed to LiDAR plus sensor redundancy 2026-04-20-autoresearch-tesla-fsd.
The divergence isn't just a sensor choice — it cascades into unit cost, operating domain, failure modes, and regulatory exposure.
Evidence
- Failure-mode exposure. nhtsa's EA26002 against Tesla specifically targets reduced-visibility conditions (sun glare, fog) where camera-only perception degrades — exactly the class of failure sensor fusion is designed to handle 2026-04-20-autoresearch-tesla-fsd.
- Scale achieved under each approach. waymo (sensor fusion) runs ~500k paid rides/week across 10 cities fully driverless; tesla-fsd (vision-only) runs ~4–8 unsupervised Model Ys in Austin with remote supervision 2026-04-20-autoresearch-tesla-fsd.
- Per-unit cost vs ODD tradeoff. Vision-only lowers per-car hardware cost at consumer scale but (per the regulators) has not demonstrated the safety case for driverless operation. Sensor-fusion stacks are more expensive per vehicle but operate driverless today 2026-04-20-autoresearch-tesla-fsd.
Design implications
- If vision-only is sufficient for L4: Tesla's fleet-learning flywheel and per-vehicle cost advantage dominate long-run.
- If vision-only is not sufficient for L4: Tesla either adds sensors (expensive retrofit / hardware reset) or caps at L2 consumer ADAS + narrow robotaxi geofences.
Contradictions / tensions
- Tesla's community-tracker disengagement rate (~809 city miles to critical disengagement on v14.2) is ~37× worse than Waymo's ~30,000-mile safety-driver-removal threshold — suggesting the vision-only stack has not yet reached sensor-fusion reliability 2026-04-20-autoresearch-tesla-fsd.
- But: Tesla claims its scale of real-world data closes the gap faster than Waymo can via simulation + HD-mapped territory. Thin evidence either way in this source.
Open questions
Related
Referenced by