Fast Technology News on November 16th, Xiaomi HAD (Xiaomi Hyper Autonomous Driving) will fully integrate end-to-end large models, marking a significant upgrade in intelligent driving experience. The pioneer version is expected to be rolled out by the end of December.

Xiaomi HAD will be available on Xiaomi SU7 Pro, Xiaomi SU7 Max, and Xiaomi SU7 Ultra.

So, what is an end-to-end large model? How does Xiaomi's end-to-end large model differ from mainstream models? Xiaomi Automobile has provided a detailed explanation.

Simply put, an end-to-end large model functions like a human brain. When you see a red light, your brain immediately tells you to stop. An end-to-end large model serves as the 'mechanical brain' of intelligent driving.

Before the advent of end-to-end technology, traditional intelligent driving systems were largely composed of modular sub-systems, akin to individual workshops that processed sensor data to output driving commands.

The advantage of a modular architecture is clear division of labor. If an issue arises, it can be isolated and resolved within the respective module. However, the disadvantage is also apparent: the upper limit of performance is low. When performing complex driving tasks, there is a noticeable mechanical feel, and sometimes manual intervention is required.

End-to-end technology, on the other hand, involves direct input from sensors at one end and output of driving commands at the other. It integrates perception, prediction, and planning into a unified large model, performing all calculations within this single model. By eliminating the need for information to flow through multiple stages, errors and delays are significantly reduced.

With just one model, sensory information from cameras, millimeter-wave radars, LiDAR, and other sensors can be converted into specific operational commands such as steering wheel angle, acceleration pedal depth, etc. Through extensive data training, this technology achieves human-like driving without the need for rule-writing or handling extreme scenarios, significantly raising the upper limit of intelligent driving systems.

Xiaomi notes that current mainstream end-to-end large models are generative models that directly produce a trajectory. If training is inadequate, these models may generate illogical 'hallucinations,' indicating a low lower limit of performance.

Xiaomi's intelligent driving system has a unique approach to end-to-end operation, shifting from trajectory imitation to trajectory screening. The large model first generates all reasonable trajectories in real-time and then evaluates each for safety, comfort, compliance, efficiency, and other factors to select the optimal solution for execution.

After fully integrating the end-to-end large model, vehicle trajectories become more human-like, smoother, and more natural, with significant reductions in abrupt acceleration and braking. Additionally, vehicles can intelligently predict bus lane restrictions, plan optimal routes, and make more reasonable driving decisions in advance.

Xiaomi's intelligent driving system, integrated with the end-to-end large model, also enables seamless ETC and toll booth passages, from parking spot to parking spot. It intelligently reduces speed before toll booths, waits flexibly, recognizes entries and exits precisely, and accelerates promptly after barriers are lifted, demonstrating high efficiency.