Jia Haonan was sent from the Vice Pilot Temple
Smart car reference | Official account AI4Auto
The most important developments in the field of smart cars today:
A new version of Tesla's FSD is online.
Why is it important?
FSD Beta 10.11 version, Musk personally confirmed that there are architectural improvements, which is a major update.
Moreover, Musk also said that because the new 10.11 capability is enough for Bull X, more FSD test places will be considered in the future.
From a larger level, the most rapid progress, pure visual route representative, data system construction is the most complete... The extent to which Tesla's AI is done is a bellwether that consumers and the driverless car industry are keeping an eye on.
Let's take a closer look at where the pure visual autonomous driving technology No.1 represented by FSD has progressed.
11 updates, why does Musk value this one the most?
The 11 updates are:
Model lane geometry from dense rasters ("point packs") to autoregressive decoders.
Improve the autonomous driving system's understanding of right-of-way (road category, drivable area) in the event of inaccurate maps or failed navigation. Modeling roads at forks in the road, in particular, now relies entirely on neural network predictions rather than map information.
The accuracy of the VRU (Vulnerable Traffic Participant) detection was improved by 44.9%. Greatly reduces false alarms for motorcycles, scooters, wheelchairs and pedestrians in rainy days, mottled roads, etc. This is achieved by increasing the amount of data for the next generation of automatic recognizers, training previously frozen network parameters, and modifying the network loss function.
Reduced the very close VRU prediction speed error by 63.6%. The measure is the introduction of a new simulated adversarial high-speed VRU dataset. This update improves autonomous driving control of fast-moving and cut-in VRU.
The vehicle's climbing posture is improved, and the acceleration or braking force is stronger at the beginning and end of the climb.
Improved static obstacle perception network, improving the perception and recognition of obstacles around the vehicle.
By increasing the dataset size by 14%, the recognition error rate of the vehicle "parked" attribute (referring to other vehicles on the road) was reduced by 17%, and the accuracy of the brake lights was also improved.
By adjusting the loss function, the vehicle's ability to drive autonomously in difficult scenarios is improved. The "passable state" speed error is improved by 5%, and the speed error in high speed conditions is improved by 10%.
Improved detection and control of open doors of vehicles on the side of the road.
Optimized body control algorithms for both lateral and longitudinal acceleration in the vehicle and bumps for a smoother cornering experience.
Ethernet data transmission optimization improves the stability of FSD Ul visualization.
All the updates can be roughly divided into two categories, the first is the improvement of the passenger riding experience.
For example, the three items of bending smoothness, UI visualization, and climbing posture.
The second type is closely related to the perception and decision-making of the vehicle itself, such as the use of neural network predictions to model intersections, reducing the dependence on high-precision maps.
In addition, at the end of last year, users frequently reflected the "ghost brake" problem, the reason is that the camera is not accurate in the case of interference.
This update improves the recognition accuracy of different targets and states by adding recognizers, enriching data sets, and adjusting loss functions in the back-end AI neural network.
The progress of these capabilities, whether it is perceptual decision-making or modeling, is inseparable from the most basic lane and target prediction capabilities.
This is also the "architecture-level" update that Musk values the most in this update:
Modeling of lane geometry was upgraded from dense rasters to autoregressive decoders.
What do you mean?
Lane grating was originally a technique often applied to toll booth ETC, which extracts contour characteristics by obscuring the light by objects and is used to distinguish between different vehicles.
Tesla's dense raster modeling here is also through the virtual dense raster to extract a large number of feature points in the image data to restore and reconstruct the digital model.
The autoregressive decoder decoder uses transformers to directly predict and connect vector space lanes point by point.
The so-called "vector space" lane refers to the decomposition of the overall situation of the lane into several key parameter information, such as width, material, color, lane line type, and so on.
Why? Because for human beings, the moment they see the scene, they can understand the basic situation.
But for neural networks, it is necessary to extract key parameters that can be "understood". If each key parameter is considered a vector, then all the information contained in a lane is a multidimensional space.
With this concept, the update at the architectural level in Musk's mouth can actually be simply understood as the modeling process eliminates the intermediate step of extracting a large number of points from the image and directly generates the parameter information that AI can understand.
The benefit of this is that the system is more efficient at predicting the behavior of the road, other targets, and the back-end fusion of multiple sensor information.
The intermediate steps are reduced, and the cost of hashing power and the error rate are correspondingly reduced.
How do you rate this update?
In essence, the previous modeling methods, relying on the extraction of features from the image data that serve the "human eye", reflected the idea of human priority.
The ultimate goal of autonomous driving is to let AI replace humans, if you also extract parameters from the information data that serves humans, isn't it superfluous?
Not to mention that the calculation cost and error rate will increase.
Therefore, this update reflects Musk's deep understanding of the nature of AI and implements the "first principles" that he has always emphasized before.
Cold, rational, on the road to abandoning the inherent thinking of human beings, Musk has gone farther and farther.
For tesla FSD products, the "machine" is more thorough. Perhaps it is currently inferior to some friends in terms of MPI and stability, but from the underlying structure, it is already a purer AI.
What does more AI Tesla mean?
Musk believes that more AI and more essential FSD represent stronger autonomous driving capabilities.
Currently, FSD 10.11 is pushed within Tesla's internal employees, but Musk said that if it "performs well," it will be pushed to a larger range of user tests.
This "performance" includes two aspects, both the ability of the algorithm and the reliability of the owner.
Yes, if you want to experience Tesla FSD, you must be an "honor student".
Tesla rated every user who applied for the FSD test to determine whether he was a qualified and responsible driver.
Dimensions include 5 aspects: forward collision warning per 1,000 miles, emergency braking, sharp turns, unsafe follow-up, and forced release of Autopilot.
That said, driving distractingly, aggressively, and distrustful of autonomous driving can all result in low scores.
At present, only 98 points or more are eligible to test FSD, and the total number is only about 60,000.
Musk's talk of scaling up is also lowering the threshold for drivers to score 95 points.
And foreign users who heard the news have been excited:
Originally, the owner of the car paid for the service, but Musk played the superiority of "merit-based admission".
To say that the user psychology, but also has to be Ma Yilong.
Are There any features that Tesla owners, or other smart car users, have the most impressive features in their experience?
—Ends—
【Smart car reference】Original content, without the authorization of the account, is prohibited to be reproduced at will.
Click here to pay attention to me, remember the star, Mo Mo Da ~