In addition to Weibo, there is also WeChat
Please pay attention
WeChat public account
Shulou
2025-04-04 Update From: SLTechnology News&Howtos shulou NAV: SLTechnology News&Howtos > IT Information >
Share
Shulou(Shulou.com)11/24 Report--
CTOnews.com August 21 news, just released Tesla software update 2022.16.3.10 (that is, fully autopilot FSD 10.69). @ ACPixel released the release notes for the new version.
At the same time, YouTuber @ Whole Mars Catalog gave a 35-minute video of a real car test without manual takeover. Let's take a look at it.
A new "deep lane guidance" module is added to the vector lane (Vector Lanes) neural network, which combines the features extracted from the video stream with rough map data, that is, the number of lanes and lane connectivity. Compared with previous models, this architecture achieves a 44% error rate on the lane topology and smoother control of the lane and its connectivity before it becomes visually obvious. This provides a way for each autopilot to drive as well as someone drives their own commute, but to adapt to road changes in a sufficiently common way.
By better simulating the system and drive delay in trajectory planning, the overall driving stability is improved without sacrificing the delay. Now, the trajectory planner independently considers the delay from the steering command to the actual steering drive, as well as the delay from the acceleration and braking command to the drive. This leads to a more accurate trajectory of the vehicle driving model. This allows better downstream controllers to track peace and stability, while also allowing more accurate responses to demanding manipulations.
When approaching and leaving the middle intersection area, in the case of high-speed traffic, the unprotected left turn is improved, and the speed curve is more suitable ("Chuck Cook style" unprotected left turn). This is achieved by allowing optimizable initial twitches to mimic the tough pedals that humans step on when they need to drive in front of high-speed objects. The lateral profile close to this safe area is also improved to allow for better posture so that it is well aligned when leaving the area. Finally, the interaction with objects that are entering or waiting in the middle crossing area is improved to better simulate their future intentions.
The control of arbitrary low-speed movement from the occupied network is added. This also gives finer control over the shape of more precise objects, which are not easily represented by cube primitives. This requires predicting the speed of each three-dimensional voxel. We can now control the slow-moving UFO.
The placeholder network has been upgraded to use video instead of images in a single time step. This time background makes the network robust to temporary blocking and can predict the occupation flow. At the same time, the basic truth is improved through semantic-driven outlier elimination, hard case mining and increasing the size of the dataset by 2.4 times.
Upgrade to a new two-phase architecture to generate object kinematics (such as velocity, acceleration, yaw rate), where network computing is assigned O (object) instead of O (space). This increases the estimated speed of distant passing vehicles by 20%, while using only 1/10 of the computation.
The stationarity of the protected right turn is improved by improving the connection between the traffic light and the taxiway and the connection between the yield sign and the taxiway. This reduces the wrong deceleration in the absence of related objects and improves the position of giving way when there is no related object.
Reduced the wrong deceleration near the crosswalk. This is based on the movements of pedestrians and cyclists to improve their understanding of their intentions.
By updating the full vector lane neural network, the geometric errors of self-correlated lanes and cross lanes are increased by 34% and 21%, respectively. By increasing the internal size of each camera feature extractor, video module, autoregressive decoder and hard attention mechanism, the fine location of the lane is greatly improved and the information bottleneck in the network architecture is eliminated.
When crawling, make the speed curve more comfortable so that you can stop more smoothly when protecting objects that may be obscured.
By doubling the size of the automatically tagged training set, the animal's memory rate was increased by 34%.
At any intersection where objects may cross their own path, with or without traffic control, they can crawl to gain visibility.
By allowing dynamic resolution in trajectory optimization, the accuracy of the stop position in key scenes with crossed objects is improved, and more attention is paid to more finely controlled areas.
By involving the topology markers in the attention operation of the autoregressive decoder and increasing the loss applied to the bifurcation marks during training, the recall rate of the bifurcated lanes is increased by 36%.
By improving the on-board trajectory estimation as input to the neural network, the speed error of pedestrians and bicycles is increased by 17%, especially when I am turning a corner.
By adjusting the loss function used during the training and improving the label quality, the recall rate of object detection is improved, and the missed detection of 26% of distant passing vehicles is eliminated.
By incorporating the yaw rate and lateral motion into the likelihood estimation, the future path prediction of the object under the condition of high yaw rate is improved. This helps objects move into or out of their own driveway, especially at intersections or cut into the scene.
By better dealing with the upcoming map speed changes, the speed of entering the highway has been increased, which increases the confidence of being integrated into the highway.
By considering the bumps of the leading vehicle, the delay of starting from the stop is reduced.
By evaluating their current motion status and expected braking curve, red light runners can be identified more quickly.
CTOnews.com learned that Tesla's FSD Beta program currently has about 100000 public testers in North America.
In his last earnings call, Musk revealed that the FSD Beta test team had traveled more than 3500 million miles, far ahead of its competitors. The number of FSD Beta testers is likely to grow further by the end of the year. Musk also said that the Beta version of Tesla for the right rudder will be released by the end of 2022.
Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.
Views: 0
*The comments in the above article only represent the author's personal views and do not represent the views and positions of this website. If you have more insights, please feel free to contribute and share.
Continue with the installation of the previous hadoop.First, install zookooper1. Decompress zookoope
"Every 5-10 years, there's a rare product, a really special, very unusual product that's the most un
© 2024 shulou.com SLNews company. All rights reserved.