Rewritten AI Compiler and Runtime
The AI compiler and runtime have been rewritten from the ground up using MLIR (Multi-Level Intermediate Representation), resulting in a 20% faster reaction time. This is a fundamental infrastructure improvement that benefits all FSD driving scenarios.
Upgraded Reinforcement Learning
The reinforcement learning stage of FSD neural network training has been upgraded, improving the system’s ability to handle complex driving scenarios through better learned behaviours.
Enhanced Vision Encoder
The vision encoder has been improved for rare and low-visibility scenarios with better 3D geometry understanding. This helps the system handle unusual situations more reliably.
Improved Animal Handling
Focused reinforcement learning training on animal scenarios improves detection and response to animals on or near the roadway.
Parking Spot Prediction
Better parking spot selection with map pin prediction. A “P” icon now displays when the system identifies suitable parking at the destination.
Additional Improvements
- Reduced lane biasing and tailgating behaviours
- Improved responses to emergency vehicles, school buses, and right-of-way violators
- Better traffic light management at complex intersections
- Improved handling of unusual roadway objects
- Better temporary system degradation tolerance with automatic recovery
- Blind spot warning while parked (Cybertruck)
Hardware
Requires HW4 (Hardware 4). Bundled with firmware 2026.2.9.6.