Perception System
Vision-Only Approach
Uses 8 cameras, no radar or LiDAR.
SDF (Signed Distance Field)
- 3D environment reconstruction
- Precision: 10cm (3× improvement over V12’s ~33cm resolution)
- Multi-frame spatiotemporal fusion for dynamic object tracking
Decision and Control
End-to-end Neural Network
Replaces approximately 300,000 lines of C++ code.
- “Black box” AI decision-making, producing emergent behaviors
- Model scale 10× larger than V13
Driving Modes
- Sloth: Most conservative
- Mad Max: Most aggressive
Hardware
- V14 requires HW4 (compute power 2× that of HW3)
- Improved video compression preserves details
- “V14 Lite” with HW3 support planned for 2026
Improvements Over V12
- Higher precision environment modeling
- Better unprotected left turns, construction zones, emergency vehicle handling
- Smoother lane changes, less hesitation
- More human-like decision-making