2021.44.30.10

Full Self-Driving Beta v10.9

• Improved intersection extents and right of way assignment by updating modeling of intersection areas from dense rasters ("bag of points") to sparse instances. Increased intersection region IOU by 4.2%. The sparse intersection network is the first model deployed with an auto-regressive architecture that runs natively with low latency on the TRIP Al accelerator chip, through innovations in the Al compiler stack.
• Upgraded generalized static object network to use 10-bit photon count streams rather than 8-bit ISP tonemapped images by adding 10-bit inference support in the Al compiler stack. Improved overall recall by 3.9% and precision by 1.7%.
• Made unprotected left turns across oncoming lanes more natural by proceeding straight into intersection while yielding, before initiating the turn.
• Improved lane preference and topology estimation by 1.2% with a network update and a new format for navigation clues.
• Improved short deadline lane changes with better modeling of necessary deceleration for maneuvers beyond the lane change.
• Improved future paths for objects not confined to lane geometry by better modelling of their kinematics.
• Made launches from stop more calm when there is an imminent slowdown nearby.
• Improved gap selection when yielding to a stream of oncoming cars on narrow roads.