– Improved decision-making framework for unprotected left turns with better modeling of object response to ego actions by adding more features that make up the go/no-go decision. This increases the robustness of noisy measurements while being more consistent in making decisions within a margin of safety. The framework also makes use of medium safe areas when necessary to maneuver through large turns and accelerate further through maneuvers when needed to safely exit an intersection.
– Improved crawl visibility using more precise lane geometry and higher accuracy blockage detection.
– Reduce instances of attempting uncomfortable turns through better integration with future object predictions during route selection.
– Upgraded layout to rely less on aisles to enable smooth maneuvering out of tight spaces.
– Increased cornering safety with cross traffic by optimizing the neural network architecture of lanes which greatly enhanced the recall and geometrical accuracy of crossing lanes.
The recall and geometric accuracy of all lane predictions have been improved by adding 180,000 videos to the training suite.
Reduced traffic control associated with false deceleration through better integration with the lane structure and improved behavior with respect to yellow lighting.
– Improved geometric accuracy of road edge and line predictions by adding a mixing/coupling layer with the generalized static obstacle network.
– Improved geometric accuracy and visibility understanding by retraining the generalized static obstacle network with improved data from the automatic naming tool and adding an additional 30,000 videos.
– Improved motorcycle recall, reduced speed error at pedestrians and nearby cyclists, and reduced main pedestrian error by adding new SIM data and auto-tagged data to the training set.
– Improved accuracy of the ‘Parked’ feature on vehicles by adding 41,000 segments to the training set. Resolved 48% of failures logged by our telemetry for 10.11.
– Improved retrieval for the detection of crossed distant objects by replenishing the data set with improved versions of neural networks used for auto-tagging which increased the quality of the data.
Improved balancing behavior when maneuvering around cars with open doors.
– Improved angular velocity and central velocity of non-VRU objects by upgrading them to expected network tasks.
– Improved comfort when changing lanes behind vehicles with extreme deceleration through tighter integration of the pilot vehicle’s future movement estimate and the planned lane change coil.
– Increased dependence on the acceleration predicted by the network for all moving objects, which were previously only longitudinally related objects.
– Updating nearby vehicle assets with visualization indicating an open vehicle door.
– The +1.8fps system frame rate has been improved by removing three old neural networks.
Hit the Record Video button in the top bar user interface to share your feedback. When pressed, your vehicle’s external cameras will share a short autopilot snapshot linked to the VIN with the Tesla engineering team to help make improvements to the FSD. You will not be able to view the clip.
Interpretation of the FSD Release Notes
#Official #Tesla #Release #Notes