Finding
Two camps: Tesla bets on cameras alone, while China made lidar a standard part and robotaxi operators fuse everything
Sensor fusion combines cameras, radar and lidar into one model of the scene. Tesla drives on cameras alone and argues that more data beats more sensors. Waymo, Zoox, Baidu and Pony.ai fuse all three, and China's automakers made lidar a mass-market part: Gaogong counted 3.24 million lidars fitted as standard on Chinese passenger cars in 2025, more than double 2024, with penetration above 20%.
Where fusion happens matters as much as which sensors are used. Early fusion combines raw data before detection, which keeps the most information but needs precise calibration and heavy compute. Late fusion combines each sensor's detections, which is simpler and isolates failures but loses detail. Most current systems use mid-level or feature fusion inside a bird's-eye-view network, where camera, radar and lidar features meet in one top-down grid. Multi-lidar setups (three to six per car) are spreading in China for Level 3 redundancy.
Fusion levels
Where fusion happens
| Level | What is combined | Strength | Cost |
|---|---|---|---|
| Early (raw data) | Combines raw sensor data before detection | Most information; best for small or partly hidden objects | Calibration and compute heavy |
| Mid-level (features) | Combines learned features, often in a bird's-eye-view grid | Balance of detail and robustness; standard in end-to-end systems | Needs large training sets |
| Late (objects) | Combines each sensor's detections | Simple; failures stay isolated | Loses detail; weaker in edge cases |
Sensors
What each sensor contributes
| Sensor | Strength | Weakness | Unit cost |
|---|---|---|---|
| Camera | Color, texture, signs, lanes | Night, glare, fog | Low |
| Radar | Range and relative speed in any weather | Low angular detail (improved by 4D imaging radar) | Low |
| Lidar | Precise 3D geometry | Heavy rain, snow; cost | Medium, falling (~$200 for mass-market units) |
Sources
Sources and scope
- Scope
- Sensor strategies and fusion methods
- Sources
- Gaogong via China Daily (April 2026); Hesai 2025 results; company disclosures
- Updated
- October 2026