Tesla has begun to push the FSD Beta V11 version of the assisted driving system to testers, and overseas users will receive the push on the 12th. Elon Musk also confirmed that Tesla will “roll out widely” to more North American users in a few weeks.
For the FSD V11 version, Elon Musk had previously pointed out at AI Day 2022 that V11 will be a single-stack system, which will make FSD Beta’s performance smoother and its judgments smarter. So most of the industry believes that FSD Beta V11 will be a major update to Tesla’s advanced driver assistance system.
Elon Musk said at the time that he himself was using a single-stack version of FSD Beta, so he was able to reach his destination without touching the vehicle controls.
From the release notes, FSD Beta V11 (2022.40.5) has been updated and optimized mainly in the following areas.
Enabling FSD Beta on the highway.
This unifies the vision and planning stack on and off the freeway, and replaces the four-year-old legacy freeway stack. The legacy freeway stack still relies on a few single cameras and a single frame network, and is set up to handle simple lane specific operations. the multi-camera video network and next generation planning system of FSD Beta allows for more complex agent interactions with fewer lane dependencies, making way for smarter behavior, smoother control and better decision making.
Improved accuracy of the Occupancy Network for near-obstacle recall and bad weather conditions, a 4x increase in transformer spatial resolution, a 20% increase in image featureer capacity, improved side camera calibration, and an additional 260,000 video training clips (both actual and simulated).
The predicted speed error for very close motorcycles, scooters, wheelchairs and pedestrians was reduced by 63.6%. To this end, we introduced a new dataset that simulates adversarial high-speed VRU interactions. This update improves the autopilot control system around fast-moving and cutting VRU’s.
The creeping profile was improved to increase the acceleration at the beginning and end.
Improved control of nearby obstacles by using a general static obstacle network to predict continuous distance from static geometry.
Reduced the error rate of the vehicle “park” attribute by 17% by increasing the dataset size by 14%, while also improving the accuracy of brake lights.
Improved performance in difficult scenes by adjusting the loss function to improve the speed error in clear scenes by 5% and in highway scenes by 10%.
Improved detection and control of open car doors.
Improved the smoothness of turns by using an optimization-based approach to determine which lane routes are independent of the computation of a given lateral and longitudinal acceleration and steep lift limits when calculating vehicle kinematics.
Optimized the Ethernet data transfer pipeline by 15%, thus improving the stability of the FSD UI visualization.
Improved recall of vehicles behind the ego and increased the accuracy of the vehicle detection network.
