The full beta update of Tesla Self-Driving v10.12 has been released

Tesla has launched a new beta update of Full Self-Driving (FSD) software with improvements based on more than 250,000 training videos from its fleet.

Based on the pre-release notes, this is a big update.

FSD Beta allows Tesla vehicles to drive autonomously to the destination entered in the car’s navigation system, but the driver needs to be vigilant and always be ready to take control.

Because the responsibility lies with the driver, not the Tesla system, it is still considered a second-tier driver assistance system, despite its name. It was like a two-step, one-step back program, with some upgrades regressing in terms of management capabilities.

Tesla often releases new software updates for the FSD Beta program and adds more owners to it.

The company now has about 100,000 owners in the program and with more people in it is expected to have more data to train their neural networks.

Today, Tesla has launched a new beta update of the FSD software (2022.12.3.10), and, according to release notes, this is one of the most extensive updates to date.

Interestingly, Tesla for the first time notes the number of videos taken from the fleet and used to teach certain new behaviors. The automaker mentioned a total of more than 250,000 new videos used in the training kit for this update.

Tesla also said it had removed three old neural networks from the system, improving the system’s frame rate by 1.8 frames per second.

The release notes also mentioned many other improvements – some related to the level of confidence at which the system takes action, which has been a source of frustration when using FSD Beta in the past.

You can read more about all the improvements in the release notes below:

Release Notes FSD BETA v10.12

• Updated decision base for unprotected left turns with better simulation of objects’ response to ego actions by adding additional features that shape the “go / don’t go” decision. This increases resistance to noisy measurements while being more sticky to solutions within the safety margin. the frame also uses medium safe regions when needed to maneuver at big turns and stronger acceleration when maneuvering when a safe exit from an intersection is required.
• Improved creep visibility using more accurate band geometry and occlusion detection with higher resolution.
• Reduce the number of attempts to make awkward turns by better integrating with objects that predict the future when choosing a lane.
• Updated scheduler to rely less on lanes to ensure smooth maneuvering with limited space.
• Improved safety of turns crossing lanes, by improving the architecture of the neural network of lanes, which greatly increased the memory and geometric accuracy of the lane intersection.
• Improved the recall and geometric accuracy of all bands by adding 180,000 videos to the training set.
• Reduce false decelerations associated with traffic control, thanks to better integration with lane structure and improved behavior with respect to yellow traffic lights.
• Improved geometric accuracy of road edge and line prediction by adding a mixing / communication layer with a generalized network of static interference.
• Improved geometric accuracy and understanding of visibility by retraining the generalized network of static interference with improved auto-label data and by adding another 30,000 videos.
• Improved motorcycle memory, reduced speed errors of pedestrians and cyclists nearby, and reduced pedestrian course error by adding new simulator data and automated data to the training kit.
• Improved the accuracy of the “parked” attribute on vehicles by adding 41,000 clips to the training set. Resolved 48% of cases of failure recorded by our telemetry from 10.11.
• Improved response to the detection of distant intersecting objects by regenerating a data set with improved versions of neural networks used in automated labeling, which has improved data quality.
• Improved compensation behavior when maneuvering around cars with open doors.
• Improved angular velocity and lane velocity for non-VRU objects by upgrading them to network predictive tasks.
• Improved lane comfort behind vehicles with a sharp deceleration due to closer integration between the assessment of the future movement of leading vehicles and the planned lane change profile.
• Increased dependence on network-predicted acceleration for all moving objects, previously only longitudinal objects.
• Updated vehicles nearby with a visualization that shows when the car door is open.
• Improved system frame rate +1.8 frames per second by removing three old neural networks.

Subscribe to Electrek on YouTube to get exclusive videos and subscribe to podcast.